{"collectionById":{"76d23fbd-a202-403b-8313-b3bc36d46679":{"id":"76d23fbd-a202-403b-8313-b3bc36d46679","name":"Research Post","fieldSchemas":[{"id":"1d8c181d-1ccd-446e-87b0-482ccb3ee240","name":"Title","type":"plain_text","role":"primary"},{"id":"571c37b5-f5fc-4742-a5de-b4da6f3415b3","name":"Slug","type":"slug","role":"slug"},{"id":"3fcb25bf-f8b6-47c4-84d6-226369594160","name":"Content (HTML)","type":"rich_text"}],"itemById":{"809bd036-e216-4bdb-ae70-dbd1d65b4804":{"id":"809bd036-e216-4bdb-ae70-dbd1d65b4804","index":"\"NO","collectionId":"76d23fbd-a202-403b-8313-b3bc36d46679","fields":[{"id":"e1a67196-e141-41ac-9451-007e1b7bc2d2","value":"Fable 5: Expensive Intelligence Needs Expensive Work","itemId":"809bd036-e216-4bdb-ae70-dbd1d65b4804","fieldSchemaId":"1d8c181d-1ccd-446e-87b0-482ccb3ee240"},{"id":"5075e0fc-abee-4c00-8a27-9c5eb1cbb832","value":"{\"root\":{\"children\":[{\"children\":[],\"direction\":\"ltr\",\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"By now, \",\"type\":\"text\",\"version\":1},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Fable 5\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"link\",\"version\":1,\"rel\":null,\"target\":\"_blank\",\"title\":null,\"url\":\"https://www.anthropic.com/news/claude-fable-5-mythos-5\"},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" is no longer just a launch story. The more interesting signal is how divided the reaction stayed after developers had time to use it.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In some workflows, Fable 5 felt closer to a project owner: it could gather context, call tools, launch sub-agents, check intermediate results, and keep moving.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In others, it looked like a smaller step from Opus 4.8, GPT-5.5, or the strongest Sonnet-class models.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"That split matters because it points to a harder market question: which work is valuable enough to pay frontier-model prices?\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Fable 5 is testing the premium end of that question, with a 1M-token context window and premium API pricing. \",\"type\":\"text\",\"version\":1},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"GLM-5.2\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"link\",\"version\":1,\"rel\":null,\"target\":\"_blank\",\"title\":null,\"url\":\"https://z.ai/blog/glm-5.2\"},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" and other lower-cost models are testing the other end, making it harder to ignore how much work may be good enough for cheaper intelligence.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"So we hosted a closed-door\",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":3,\"mode\":\"normal\",\"style\":\"\",\"text\":\" Best Ideas \",\"type\":\"text\",\"version\":1},{\"detail\":0,\"format\":1,\"mode\":\"normal\",\"style\":\"\",\"text\":\"discussion with builders, researchers, and founders testing these systems in real workflows.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":1,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The market structure we kept coming back to was simple: frontier models may capture 80% of model revenue, while lower-cost and open-source models may process 80% of tokens. The real question is how thick the top of the task pyramid is: how much work is valuable enough, difficult enough, and measurable enough to justify frontier prices.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Where Fable 5 Actually Shows Its Edge\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The first signal was not a simple thumbs-up or thumbs-down. Builders were describing different task conditions:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Some users saw a real step forward. Fable 5 started to feel more like a project owner. It could notice missing context, decide what needed to be done, call tools, assign subtasks, inspect intermediate results, and keep a project moving with less hand-holding.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Other users came away less convinced. Compared with Opus 4.7 or 4.8, Fable 5 felt steadier in some places and more expensive in others, with no obvious jump in underlying intelligence.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Even inside coding, the reaction varied by task. Some users found Fable especially strong on SQL-like work. In Python-heavy work, it felt much closer to GPT-5.5.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Why the Same Model Can Feel Different\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The split makes more sense once the examples are sorted by what the task asks the model to do.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Personal planning, document formatting, routine coding, and ordinary chat are poor tests of frontier capability. Current models already do well there. The acceptance criteria are often fuzzy. A slightly better answer is rarely worth a much higher price.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Complex engineering, reverse engineering, auto-research, security analysis, large system refactoring, and workflow construction create a different test. These tasks force the model to gather hidden context, reason over many steps, recover from failures, make tools, inspect environments, and choose the next action. In those settings, Fable 5 starts to separate from models that already feel good enough on everyday work.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Research work in AI labs made the difference easier to feel.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"One AI researcher in the discussion said that in some open research directions, Mythos-class models are getting close to human researcher judgment. The difference became obvious when switching from Claude back to Gemini 3.5 in the same research direction. Work that felt roughly 70% complete with Claude could fall back toward 50% after the switch because of hallucinations, unstable instruction following, and the model’s inability to keep following the planned experimental path.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Task type was not the only reason people disagreed. The result also depended on which Fable 5 path a user actually encountered:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Some users turned Fable 5 down to low effort and found that it felt more like a stable Opus 4.6. The quality remained usable while token usage dropped sharply.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Others saw simpler sub-agent steps delegated to cheaper models such as Haiku.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In \",\"type\":\"text\",\"version\":1},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"safety-sensitive categories\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"link\",\"version\":1,\"rel\":null,\"target\":\"_blank\",\"title\":null,\"url\":\"https://www.anthropic.com/legal/aup\"},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\", requests can trigger refusal, constraint, or fallback to Opus 4.8 or another model. Cybersecurity, biochemistry, model distillation, and certain AI-for-science use cases are especially likely to hit these boundaries.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"type\":\"image\",\"version\":1,\"hash\":\"6397d41357b7a5275820ebed97b55df4fa8d6d4a\",\"src\":\"https://substack-post-media.s3.amazonaws.com/public/images/1e677eec-ed82-4394-87dd-c4de16df8491_1542x1028.png\",\"altText\":\"\",\"originalImageWidth\":1542,\"originalImageHeight\":1028,\"isFillWidth\":false}],\"direction\":\"ltr\",\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":2,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Source: Reddit discussion screenshot\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":2,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In other words, two developers can both be accurately describing Fable 5 while seeing different effort settings, routing paths, safety behavior, and cost structures.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"That task-level split also shaped the investment read.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"On the upside, Fable 5 makes high-value work, super-engineering, and auto-research materially closer to usable. Model progress is still moving.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The more cautious read is that Fable 5 has not yet proved it can expand the market enough. The model is powerful, the usage threshold is high, and the market still has no clear answer on how much high-value work exists.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"That turns the investment question into a task-level question: not simply whether Fable 5 is expensive, but whether enough work is valuable enough for that expense to make sense.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The Price Only Matters Per Task\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Fable 5’s token price is roughly twice Opus 4.8: USD 10 per million input tokens and USD 50 per million output tokens. In many cases, that is expensive.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"type\":\"image\",\"version\":1,\"hash\":\"35186c8ff2f76df577bf247d550bde4663dfdf83\",\"src\":\"https://substack-post-media.s3.amazonaws.com/public/images/a56f369c-e8eb-4a36-a60d-edfe964cf458_1252x522.png\",\"altText\":\"\",\"originalImageWidth\":1252,\"originalImageHeight\":522,\"isFillWidth\":false}],\"direction\":\"ltr\",\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":2,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Source: Anthropic API pricing page\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":2,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Task cost, however, depends on more than token price. If Fable uses fewer output tokens and less intermediate reasoning, total usage can fall. In some tasks, Fable’s token usage may be about half of Opus 4.8. In low effort mode, intermediate usage can fall below one fifth of the higher-effort path. Under those conditions, low-effort Fable can cost less end to end than Opus 4.8, despite the higher token price.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"CursorBench 3.1\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"link\",\"version\":1,\"rel\":null,\"target\":\"_blank\",\"title\":null,\"url\":\"https://benchlm.ai/benchmarks/cursorBench31\"},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" shows the other side of the same curve. In that benchmark, which compares the 75% score threshold against average task cost, Fable 5 Max reportedly achieved the highest score and the highest cost per task. Test-time compute scaling still works, although the curve is flattening. More reasoning budget still buys performance. The marginal score gain per dollar is falling.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"type\":\"image\",\"version\":1,\"hash\":\"958a9ebf505f642f5f101ec32809c4a85905fc8c\",\"src\":\"https://substack-post-media.s3.amazonaws.com/public/images/70b91c97-4c79-4711-8d1f-29f83b65348f_1068x1280.bin\",\"altText\":\"\",\"originalImageWidth\":1068,\"originalImageHeight\":1280,\"isFillWidth\":false}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":2,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Source: BenchLM, CursorBench 3.1\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":2,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The price question has to be asked at the task level: routine work can make Fable 5 look like an expensive purchase, while high-value work can make the same model look cheap.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Once the premium is measured per task, the next question is product design. The model has to sit inside a workflow where its extra capability can reach the actual bottleneck.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Enterprise Workflows Are the Real Use Case\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"If Fable’s premium only appears inside high-value tasks, enterprise adoption is less a model-choice problem than a workflow problem. The question is whether the work environment lets the model reach the bottleneck: goals, context, tools, permissions, feedback, and a clear definition of done.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The teams that felt the strongest change were not just swapping a better model into an old process. They were changing the unit of delegation. Humans set the goal, boundary, and acceptance criteria; the model decomposes the work and pushes execution forward.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"One founder described the company as starting to look like a file system: every person is an individual contributor who hands work to models, reviews intermediate output, and keeps the broader state organized.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"That metaphor is useful because it also shows what is missing. A file system needs names, folders, permissions, versioning, and recovery paths. An enterprise agent workflow needs the same kind of structure.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Fable 5 can fill in context, break down tasks, launch sub-agents, self-check, and revise. But raw agency is not the same as a reusable workflow. A model that can launch 100 sub-agents can also amplify an unclear plan 100 times. Durable enterprise workflows still need stable inputs, outputs, state management, exception handling, permissions, and acceptance standards.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"This is the practical reason teams can still spend one or two weeks cleaning up and redesigning a workflow from scratch, even when the underlying model is Opus 4.8 or Fable 5.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The useful pattern from the discussion was a sequence that starts below the model:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"type\":\"image\",\"version\":1,\"hash\":\"49b2143e8f47d12aa65920df47220209133d55d7\",\"src\":\"https://substack-post-media.s3.amazonaws.com/public/images/8f7bda21-c839-46e4-a981-06da7849a52c_2500x2003.png\",\"altText\":\"\",\"originalImageWidth\":2500,\"originalImageHeight\":2003,\"isFillWidth\":false}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The deeper enterprise AI transformation starts after that sequence is in place. It is not about replacing the existing process one task at a time. It is about redesigning the process so information moves through fewer handoffs, fewer coordination layers, shorter decision chains, and more organizational knowledge written in a form an agent can use.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"This creates a product split between coder and non-coder agents:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Coders need open, programmable systems: repositories, sandboxes, logs, shells, tests, and configuration.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Non-coders need stable task interfaces, usually GUI-based, wrapped around predefined workflows.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"For coders, an open programmable surface is the workflow. For everyone else, the workflow has to be packaged into a stable interface. Most users need a usable work surface before they need a build-your-own-agent kit; the harness is what turns raw model capability into team-level output.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The use cases below are where that workflow argument became concrete. The strongest Fable 5 signals in the discussion came from tasks near ordinary coding, but not identical to it: design, reverse engineering, complex proof-of-concept work, and long-horizon engineering.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Design and 3D Exposed a Different Kind of Taste\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In design, three.js, 3D world generation, and one-shot website creation, users reported a noticeable improvement. Fable 5 could generate more complex 3D demos, moving from simple toy games toward Minecraft-like interactive scenes. In one-shot web design, the first version often had more visual coherence. The default AI color palette showed up less often; the design felt cleaner and closer to intentional product work. The model also seemed better at stopping before the design became overworked.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Reverse Engineering Rewards Hidden Context\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Reverse engineering tested a different skill: hidden context recovery. These tasks ask the model to read a web page, obfuscated JavaScript, an Android app, a game ROM, or runtime behavior, then recover the product feature or game mechanic.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"One example involved porting a web game into Godot. Fable 5 could read the obfuscated JavaScript, recover the first-level logic, produce runnable code, and reproduce the main elements. Its weakness was visual exactness: element size, overlap, scaling, and reference-image alignment remained unstable. Even with a reference image, the model struggled to match every visual detail.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Some developers believe these capabilities began emerging around Opus 4.6, partly because safety-relevant and hacking-adjacent training signals were absorbed by the model. That raises the bar for defensive cybersecurity.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The TypeScript Eval Was a Real Frontier Task\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The task was effectively to build an Excel-like engine from a specification with thousands of lines. Fable 5 completed it in about 3 hours at a cost of around USD 200, becoming the first model to finish the task in that user’s testing. GPT-5.5 xhigh took about 10 hours and reached nearly 90% functional completion, with weaker engineering decisions and weaker performance.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Even there, Fable 5 is not the default tool for every coding task. The same user expected to keep using Codex for ordinary development. Its value is clearest when older models get stuck on a complex proof of concept or high-value task.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Long-Horizon Work Still Breaks in Human Ways\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Long-horizon work remained mixed.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The first example showed the ceiling. One CTO-level user had written more than 1 million lines of AI-assisted code over the past year. Over the most recent 3 months, his daily AI coding cost exceeded USD 1,000. After Fable 5 launched, he reran or extended prior production tasks and spent more than USD 10,000 over 2 days. His conclusion was sober: compared with Opus 4.7 or 4.8, Fable 5 mainly improved tool-use fluency and stability. The underlying intelligence felt familiar.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The second example exposed a different failure mode. Another developer had built an upper-layer plugin system around OpenClaw and deployed about 300 instances inside enterprises. The desired task was a daily auto-upgrade workflow: find the latest version, upgrade OpenClaw and Hermes, preserve internal patches, resolve upstream merge conflicts, and keep the process reusable. Starting from an existing 1.0 base, Fable 5 could handle concrete version issues and understand parts of the upgrade and patch-repair work. It repeatedly turned “write a general upgrade workflow” into “fix a patch for the June 6 version.”\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Memory was the other long-horizon limit. After a task runs for a long time, the model can still lose earlier information. Fable’s long-horizon ability has improved, with plenty of distance left from infinite context and fully reliable execution.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Taken together, these workflow examples sharpen the market question: if the frontier only shows up in certain tasks, how much revenue can those tasks support?\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Capability Splits Become Revenue Splits\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The Premium Lives at the Top of the Task Pyramid\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The split around Fable 5 turns into an economic question once a task is specific enough to route. Users will increasingly ask whether this job is valuable, difficult, and measurable enough to justify using the strongest model available.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"For a Mythos-class model, the comparison set includes every model that might be good enough: Sonnet, Opus, GPT-5.5, Gemini, DeepSeek, or a smaller local model.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"If a cheaper model can finish the work reliably, Fable’s premium is fragile. When the task sits beyond both the user and the cheaper model, the premium becomes easier to defend.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"That gives us a practical test for high-value work. It usually has three attributes:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The model has to actively acquire context; the prompt does not contain enough information.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The result can be clearly evaluated; someone can tell whether the work passed.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The task can absorb exploration cost; success is valuable enough to justify extra tokens, extra time, and failed attempts.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Those traits explain why everyday prompts often flatten the gap. Ordinary coding, life planning, casual chat, and fuzzy deliverables are often too easy, too vague, or too hard to evaluate. The frontier may be doing more work, but the user has no clean way to see or price the difference.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"ROI changes when the stakes are large enough. Token price still matters, but it is only one input. In a USD 100 million investment decision, or a core infrastructure change supporting a product with hundreds of millions of DAU, the extra token cost matters less than judgment quality, reliability, and the ability to recover from mistakes. We may still hesitate to hand those jobs to Fable or Mythos end to end, but the direction is visible.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"As costs fall, the top of the pyramid can thicken. If model costs fall by another two orders of magnitude over the next 1 to 2 years, many high-value tasks that look marginal today become worth running. At the same time, ordinary knowledge work and consumer scenarios may already be well served by Opus 4.6, Chinese models, or other low-cost models. The frontier business question is therefore a pyramid question: how much work sits at the top, how much revenue can it generate, and can that revenue support continued frontier scaling?\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"That is why model value may concentrate more sharply than human labor value. A top engineer, researcher, or investor can only work on a limited number of problems at a time. A model improvement can be packaged into an API and deployed across many workflows. Once the product wrapper is good enough, a small quality gap can be monetized across far more tasks than a human quality gap.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"A rough way to express that market is P times Q. P is the value of each task. Q is the number of tasks. Falling model costs increase Q by making more tasks worth running, while frontier capability increases P by making harder tasks feasible.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"This is why enterprise productivity may be much larger than consumer productivity. Recommendation systems, search systems, core infrastructure, internal enterprise toolchains, and large engineering refactors all consume enormous tokens. Fable 5’s multi-agent and sub-agent capabilities make super-engineering more plausible: a future model may be asked to build a Google, build an Office, or refactor an entire large system on its own.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"That still sounds like a future product. It is already a live capital-market question.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Frontier Models Take Revenue, Lower-Cost Models Take Tokens\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The same pyramid explains why cheap models matter. A workflow may need a frontier model for planning, diagnosis, or final judgment, while most of its steps only need a cheaper model that can follow instructions reliably.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"A Fable-style workflow makes that structure easier to see. The strongest model does not need to run every step. It can plan, diagnose, inspect results, and make judgment calls, while simpler subtasks go to Haiku, DeepSeek, Gemini, or local models.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"This is where the product problem begins. In Codex, Claude Code, and similar environments, users can already connect a wider model pool. The weak point is visibility: they often cannot see which model handled which step, why that model was chosen, whether quality held, or whether the routing saved money.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Low-cost model demand comes from two places at once: consumer products with enormous token volume, and enterprise workflows where a long-horizon project can be broken into many medium- or low-difficulty steps. Once work is decomposed, many steps only need GPT-4-level intelligence, not Fable-level judgment.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"That is why the low-cost layer can grow even if frontier models keep getting stronger. One developer working on enterprise scenarios said his team’s internal benchmark suggests DeepSeek V4 may already be approaching GPT-4o mini level. If a cheap model can reliably deliver GPT-4-class intelligence, it can absorb a large share of enterprise task volume.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Frontier labs can distill strong small models too, but compute scarcity changes the incentive. When high-paying customers are already supply-constrained, a frontier lab may prefer to spend scarce capacity on the top of the pyramid instead of fighting every low-price route.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"That creates room for companies optimized around cost-performance rather than maximum frontier capability. Google is the obvious example: it can coordinate chips, infrastructure, models, and product distribution, then push Flash-style routes down the cost curve.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Chinese model providers enter the same structure from a different angle. Semiconductor constraints and self-reliance make low-cost inference a strategic priority, and the commercial opening has two parts:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Later-moving model companies can learn from frontier systems that came first, then compete through distillation, compression, and cheaper inference.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Large parts of consumer and enterprise usage only need good-enough intelligence; when a task does not need a frontier model, cost, latency, and availability start to matter more.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The sharper question is whether that cost advantage can last. If frontier labs decide to optimize small models at the same intelligence level, Chinese providers will need an advantage in inference infrastructure, distribution, or use-case fit, not only cheaper model weights.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The Routing Layer Becomes the Business\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Once revenue and token volume split, the next layer of value is routing. The system has to decide which piece of work deserves frontier intelligence, which piece can go to a cheaper model, and when it should stop or ask for help. That opportunity likely sits above the model API, in the harness layer.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Coding is a special case because the loop is easy to observe. A coding agent can see the repository, run in a sandbox, inspect logs, execute tests, and verify whether a change worked. Many vertical agents only see what context went into one LLM API call. They lack the task’s ROI, definition of done, and business value.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Real routing means understanding the objective, context, acceptance standard, user preference, cost constraint, and available model pool. It means knowing when to spend Fable-level intelligence, when to call a cheaper model, when to ask the user for clarification, and when to stop.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Routing becomes the operating system for intelligence allocation.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The product that observes the full loop will be in the best position. It can see what the user meant, what the model tried, how much it cost, whether the result passed tests, and whether the user accepted it. Over time, that data can become a routing advantage.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Evals Decide What Models Become\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Evals Are Model Strategy\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Once routing decides where intelligence is spent, evals decide what kind of intelligence gets improved. They turn product taste into model behavior.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"This is one reason Fable 5 feels different in messy work. It appears stronger at implicit intent: when to change code, when to leave code alone, when to refuse a dangerous request, and how to infer the user’s real goal from incomplete context. That ability likely comes from analyzing, cleaning, and retraining on large volumes of user traces and execution trajectories.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"By the Fable 5 generation, this makes evals more than a back-office measurement tool. They are a product roadmap.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"It also explains why daily use can feel flat even when the model has improved. The improvement may live in a task domain the user has not learned to ask for, or inside an environment the user has not built yet.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"For a model company, choosing an eval means choosing what the next model will become good at. If the eval rewards direct answers, the model learns that. If it rewards long-horizon tool use, the model learns that. If it rewards intent reading, refusal judgment, and self-correction, the model learns that.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"That choice will not be the same across markets. Consumer AI and enterprise AI are likely to optimize toward different definitions of success:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Consumer AI needs a business model around user experience: fast, good, and cheap. Inference cost has to approach the cost structure of major consumer apps such as WeChat, Douyin, or TikTok before AI can enter everyday life at massive scale. Work productivity is only one part of that market. Companionship, entertainment, learning, and personal service matter too.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Enterprise AI can keep paying for higher intelligence. Large engineering tasks, high-value business processes, core infrastructure, and decision support can justify premium pricing. Reliability, controllability, and capability remain central.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Over time, the split may become less about consumer versus enterprise and more about human versus agent. Robots and software agents may become major intelligence endpoints of their own.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The same split in optimization targets shows up in lab strategy. Model companies are also training for different end states.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"For example, Anthropic and OpenAI were framed as having different strengths. Anthropic appears more committed to scaling, text- and code-based data, enterprise use cases, and large model size. OpenAI appears stronger in RL, product surface area, iteration speed, and commercial execution.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"That difference matters for Fable 5. Anthropic’s harness makes the model better at organizing work, calling tools, maintaining context, assigning sub-agents, and merging intermediate results back into the main agent. GPT-5.5 has not yet shown the same dynamic workflow training in the same way. If OpenAI closes the multi-agent and workflow gap, the gap with Fable 5 on multi-agent tasks could narrow.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The Bottleneck Has Moved to Requirements\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"After evals decide what the model is trained to become, requirements decide what the model can actually act on.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In the ideal version, a model infers what the user really wants, then decides whether it can complete the work. In practice, many users cannot clearly say what they want. The model has to infer intent, working style, tool preferences, project background, and what done means.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Requirement definition is becoming a new bottleneck, and a lot of token waste starts there. One sentence is usually not enough to complete a complex task, for the same reason a human team cannot deliver a complex project from one vague executive request.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"As tasks become more complex, requirements need structure: background, objective, boundary, definition of done, and a way to iterate. OPP, or Objective - Problem - Proposal, is one way to turn demand alignment into an engineering practice.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"This kind of structure is what turns model capability into repeatable work, whether the collaborators are humans and agents or agents working with other agents.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Some teams found that GPT-5.5 can run for many hours in real codebases when the task is broken down clearly and paired with conventions, AGENTS.md, and a definition of done.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"One project involved migrating more than 3,000 mock-dependent tests into real-environment tests. Once the goal was defined clearly as turning mock tests into real tests, GPT-5.5 completed most of the migration and improved deployment controllability.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"A large-file refactoring task showed the opposite pattern. When the team let GPT-5.5 freely refactor, it changed a small piece and stopped. When the team added a hard constraint, such as reducing files below 1,000 or 2,000 lines, the model kept working for almost two days and split 8 to 9 large files into smaller modules.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Tokenmaxxing Needs to Become Valuemaxxing\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Counting Tokens Is the Wrong Scoreboard\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"If requirements determine whether tokens are useful, then token usage alone becomes a bad scoreboard.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"If enterprises treat token consumption as the leaderboard for AI transformation, teams will reward-hack it. They will burn tokens to prove they are using AI seriously.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The better KPI is valuemaxxing: which people, practice groups, workflows, and tasks consumed tokens, and did those tokens move toward higher-value work?\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The dashboard problem is familiar from the database and Snowflake era. More queries could mean a bigger bill without more business value. Enterprises needed dashboards to understand where queries went and whether they supported real decisions.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"AI dashboards are harder because a model call can mean many things: a decision, a workflow step, a customer interaction, a coding task, or nothing useful at all. In many deployments, an FDE or field engineering team may need to help customers build the right dashboard.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The same visibility problem also puts pressure on pricing. The gap between ordinary users and heavy users is too large for simple subscription plans to hold forever. Heavy users can burn through fixed monthly plans quickly, while ordinary users resist paying for high ceilings. Codex and Claude Code are already feeling token-cost pressure. Future pricing may involve more task-level pricing, quota controls, model routing, enterprise budget pools, and outcome-based pricing.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"And before a dashboard can prove ROI, it has to answer a simpler question: where did the spend go?\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Many vertical AI vendors cannot know exactly how much value an output created for the customer. Only the customer can know that. A dashboard does not need to solve value attribution perfectly on day one. It needs to show whether tokens are flowing into more valuable workflows, so budget expansion becomes easier to accept.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"If token value becomes visible, AI growth can become healthier. It may even speed up again.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"China Will Hit Token Pressure on Its Own Curve\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"China has not yet felt the same token pressure at full intensity. The curve is different for three reasons:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The customers with the most money have not fully scaled token consumption.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In e-commerce, recruiting, advertising, and similar domains, token use can create clearer value feedback loops.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Token costs are also lower in China.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The pressure will arrive later, on a different cost curve.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"RSI: The High-Value Task Every AI Lab Is Chasing\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Closed Loops Make Token Value Visible\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"This is where the token-spend discussion turns inward. Customer-facing AI workflows still make value attribution messy: the model, the human, the toolchain, and the organization all contribute. A frontier lab has a cleaner loop. If a model helps produce better experiments, better evals, or better training directions, the value feeds directly into the next model.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"That loop is \",\"type\":\"text\",\"version\":1},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Recursive Self Improvement\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"link\",\"version\":1,\"rel\":null,\"target\":\"_blank\",\"title\":null,\"url\":\"https://www.anthropic.com/institute/recursive-self-improvement\"},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\", or RSI. It may be the highest-value internal workflow for frontier labs: using models to accelerate research, training, evaluation, and direction-setting for the next generation.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In practice, RSI looks less like a model magically rewriting itself and more like closed-loop research: literature review, hypothesis generation, experiment design, code writing, eval construction, result analysis, failure diagnosis, and next-step selection.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The path toward RSI can be broken into several stages: humans build early models, chatbots assist work, coding agents write and edit code, autonomous agents assign work, and eventually agents help build and train models. The important shift is from assistance to closed-loop model development.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"type\":\"image\",\"version\":1,\"hash\":\"e705d8f56a78f6b8484c14b6a62f8bfb837d1d3e\",\"src\":\"https://substack-post-media.s3.amazonaws.com/public/images/b2d0ff7f-85a0-4d1b-bb78-8eaf34fd7662_1880x1262.bin\",\"altText\":\"\",\"originalImageWidth\":1880,\"originalImageHeight\":1262,\"isFillWidth\":false}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":2,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Source: Anthropic, Recursive Self-Improvement\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":2,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"At this point, tokenmaxxing and RSI meet. A closed loop makes it easier to see whether tokens produced better hypotheses, better experiments, better evals, or better next-step decisions.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In ordinary human-agent collaboration, that attribution is much harder. It is difficult to isolate how much value came from the human, the model, the tool, or the organization. Tasks with clean evaluation can be optimized through RSI or closed-loop evals. Tasks without clean evaluation need benchmarks and KPIs before token spend can be judged.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"RSI Changes the Lab Race and the Org Chart\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Nowadays, top labs are already moving toward closed-loop research and auto-research. If an internal model stays ahead of public models and is used first for internal research automation, a flywheel appears: better models accelerate research, and faster research produces better models. The strongest model may serve internal R\u0026D before customers ever see it.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Once that flywheel turns, it can widen the gap between labs. It also explains why researchers inside labs may be more excited about Mythos or Fable than ordinary developers are. They are using the model on exactly the high-value tasks where frontier capability shows up.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The race can develop in two ways:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"One lab may first improve its own model-iteration speed and build a meaningful lead.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"All frontier labs may use RSI to solve harder research problems, while progress slows because of experimental resource limits and diminishing returns.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"RSI also changes organizational design. When models can run research or engineering tasks end to end, the human bottleneck shifts from execution to goal definition and evaluation. The key work becomes writing the loop, goal, and eval rubric clearly enough that the agent knows what completion means and what value means in that task.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The Next Bottleneck Is Infrastructure\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Sub-Agents Are Memory Scaling in Practice\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"When high-value tasks move into closed loops, the bottleneck shifts from the model itself to the system that feeds it context, memory, tools, and compute.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Sub-agents show the shift most clearly.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"They are not just a product feature. They change the training problem first, then the infrastructure problem.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"At the model level, there are two training paths:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The simpler path is end-to-end training of the same checkpoint inside one environment. The same model acts as both main agent and sub-agent. The main agent leans toward planning, while the sub-agent leans toward execution. One way to describe it is two objectives sharing one model: planning pressure on the main-agent role and execution pressure on the sub-agent role.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The more complex path is a large main agent coordinating one or more smaller sub-agent models. That turns the problem from training one model into training and aligning multiple models at once. If the small model learns only execution, it may lose necessary planning ability. That requires additional task design, data cleaning, and evaluation.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"At the system level, the same design can be read as memory scaling.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The main agent does not need to keep every detail inside its own context window. It can assign parts of the work to multiple sub-agents. Each sub-agent handles information in its own context and returns only the conclusion, code change, or key result.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The immediate payoff is slower consumption of the main context window and less pressure to keep expanding a single context window to 1 million tokens and beyond.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The cost shifts into shared context. When a main agent assigns work to many sub-agents, it often repeats the same user background, project constraints, tool preferences, and task objective. If those can be reused as a shared prefix, prefix caching, prompt caching, and batch inference can reduce KV cache transfer costs. One estimate was that a 50K-token KV cache reused as a shared prefix could reduce KV cache transfer overhead by about 20x in batch inference.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Sub-agents can also raise batch size and improve inference margins. In ordinary use, one user may have a limited number of parallel threads, with batch size sometimes around 34 to 35. In an agent workflow, one job can launch dozens of sub-agents and keep batch size high for longer.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"For model companies, higher batch size lowers unit inference cost. For users, it appears as stronger parallel execution, while the real token and infrastructure cost becomes harder to perceive.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Efficiency gains, however, do not guarantee lower total consumption. The catch is Jevons paradox. Storage moves in two directions:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Unit storage cost may fall. More sandbox runtime can sit on NAND. Higher utilization can reduce per-task storage cost. Repeated access to shared state across threads and sub-agents can make CXL memory pools more useful.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Total storage demand may still rise. High-value tasks will keep increasing storage and memory consumption. Hardware supply grows more linearly than intelligence demand. If workflow makes each task cheaper or more effective, users will run more tasks, increasing total storage and memory demand.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In practical terms, making a resource easier to use can increase total consumption instead of reducing it.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Compute Supply Sets the Price War Ceiling\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h3\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Memory is one limit. Compute supply is the next one.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Compute supply determines the near-term revenue ceiling of model companies. It also determines whether they can fight price wars.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"That is why any price-war discussion has to start with capacity. Some estimates suggest OpenAI has both more incentive and more room than Anthropic to cut prices for enterprise ARR and market share. \",\"type\":\"text\",\"version\":1},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Rumored GPT-5.6 price cuts\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"link\",\"version\":1,\"rel\":null,\"target\":\"_blank\",\"title\":null,\"url\":\"https://www.wsj.com/tech/ai/openai-considers-drastic-price-cuts-anticipating-war-for-users-with-anthropic-9b8c178e?eafs_enabled=false\"},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" would fit that logic.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"type\":\"image\",\"version\":1,\"hash\":\"8fd3a146fc20995a4d11564ea0c198a80fb4a9a3\",\"src\":\"https://substack-post-media.s3.amazonaws.com/public/images/4fe7423a-b9fa-47f6-8aba-1a3cf7d6a91d_1386x1582.bin\",\"altText\":\"\",\"originalImageWidth\":1386,\"originalImageHeight\":1582,\"isFillWidth\":false}],\"direction\":\"ltr\",\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":2,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Source: The Wall Street Journal\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":2,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The estimates are rough, but useful:\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Anthropic appears compute-constrained. Some market estimates put Anthropic’s compute scale around 2 to 3GW, with roughly half used for training and even tighter real supply for inference.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"OpenAI may have more compute headroom, with some estimates above 5GW. Its pricing room could come from inference efficiency improvements from B-series hardware and lower unit inference cost if OpenAI’s activated model size is smaller than Fable’s.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The long-term trend may be that flagship models get cheaper. Reasoning tokens and parameter counts can rise while API prices still fall. \",\"type\":\"text\",\"version\":1},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"GPT-4o\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"link\",\"version\":1,\"rel\":null,\"target\":\"_blank\",\"title\":null,\"url\":\"https://openai.com/index/hello-gpt-4o/\"},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\", for example, launched with API pricing roughly 50% lower than GPT-4 Turbo. Anthropic has historically cut prices less frequently.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"For Anthropic, the near-term profitability question is whether compute buying matches revenue growth. If it buys too little compute, model supply constrains enterprise ARR. If it buys too much, depreciation or lease costs raise near-term losses.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In the long run, a profitable frontier model business can still work. If leading labs can reach USD 300 billion in ARR, then even USD 30 to 50 billion of annual training cost could be rational if inference gross margin holds around 60%.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The harder variable is how much revenue each gigawatt of compute can actually support. Market estimates still vary widely, roughly USD 25 billion to USD 45 billion per GW, which means the same compute plan can look disciplined or aggressive depending on utilization, pricing, and inference margins.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Fable and Mythos access limits may also increase enterprise demand for backup models and hardware. If companies worry that frontier model access could be restricted by regulation, safety policy, or supply constraints, they will invest in alternatives: internal models, open-source fine-tuning, model reserves, and multi-vendor deployments.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Even without the newest frontier model, demand for the Opus 4.6 to 4.8 generation is already strong. The hardware supply chain remains a bottleneck.\",\"type\":\"text\",\"version\":1}],\"direction\":\"ltr\",\"format\":\"start\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"}],\"direction\":\"ltr\",\"format\":\"\",\"indent\":0,\"type\":\"root\",\"version\":1}}","itemId":"809bd036-e216-4bdb-ae70-dbd1d65b4804","fieldSchemaId":"3fcb25bf-f8b6-47c4-84d6-226369594160"}]}}}},"slugByItemId":{"d495e0e3-e929-4163-9d3e-0fec19ef9c2c":"beyond-deepseek-what-chinas-model","a54b6460-7b13-40d1-b52f-59d326989a10":"openclaw-is-the-signal-our-thesis-on-long-horzion-agents","61387d11-a584-4b82-b223-7fd449542135":"the-ai-bubble-reckoning-1999-all","f650e52a-191a-4241-a8b4-12705cb62385":"ai-for-life-science-landscape","3db0c5d9-c44c-4950-a175-45c033132ebe":"ai-coding-landscape-how-agents-disrupt","c4aeef83-34d4-411e-b40c-b432dd225928":"continual-learning-next-paradigm","809bd036-e216-4bdb-ae70-dbd1d65b4804":"fable-5-expensive-intelligence-needs-expensive-work","c23fe0ec-cf33-40e2-b090-b73d06099e9a":"generalist-and-the-270000-hour-advantage","80ff4031-030f-4caf-b7a9-1825fd8cb70f":"agi-2026-are-we-the-final-white-collar","940cd5e4-2e54-4b7a-941f-c0ea4e37df50":"beyond-the-cloud-are-llms-the-new","0c9ed8bc-cb92-4008-9f50-15867c8349d3":"our-thoughts-on-llm-part-one","fb158125-2004-4c70-972d-88d6846176ab":"decode-the-buzzword-why-harness-engineering-matters-now","cd412711-7095-43ca-8389-8011dab2f900":"inventing-anthropic-two-ingredients-behind-the-ai-winner","7b886fb4-c171-49aa-877e-475cf8b97f18":"how-openai-could-turn-the-tables","598cfcc0-365f-40c6-9f63-6631c6f5a6ff":"will-chinese-ai-leap-ahead-or-follow","91e6e6f8-7396-4490-b421-e7e2ff664063":"rl-scaling-from-research-trick-to","d285731a-bce6-45c3-9dac-26edd7829535":"pulse-how-openai-starts-outrunning"}}