Fable 5 debuts, then vanishes — and tool transparency deepens

A model week turned upside down

Monday brought what looked like the biggest launch of the quarter. Anthropic released Claude Fable 5 and Mythos 5 — both with a 1M-token context window by default, 128k max output tokens, and always-on adaptive thinking that can no longer be disabled. A new tokenizer produces roughly 30% more tokens for the same text, and safety classifiers now run on every request and during response generation, with an opt-in fallback that re-runs refused requests on another model. By Friday, the picture had flipped entirely. The US government issued an export-control directive suspending all access to Fable 5 and Mythos 5, and Anthropic confirmed access was effectively suspended. Meanwhile, older model versions were retired: Claude Sonnet 4 and Opus 4 were taken offline, and OpenAI also phased out GPT-5.2 models in ChatGPT, auto-migrating conversations to GPT-5.5.

Tool transparency and tighter agent controls

While model access shifted abruptly, the platform tools powering agentic workflows saw quieter but important polish. The code execution tool now explicitly advertises its 90-second per-cell time limit in the tool description, giving Claude the information it needs to budget long-running cells. The web search and web fetch tools gained a response_inclusion parameter that lets agentic workflows drop consumed result blocks from the API response, reducing wire noise for multi-step search loops.

For self-hosted sandbox builders on AWS, a new GET /v1/environments/{id}/work endpoint lets infrastructure poll pending work items, authorized through the GetEnvironment IAM action.

Claude Code v2.1.172 crossed a practical threshold: sub-agents can now spawn their own sub-agents, up to five levels deep. Subsequent patches fixed Fable 5 model name normalization and added an enforceAvailableModels managed setting that constrains the default model to the admin allowlist.

OpenAI ecosystem: partner network and user-visible controls

OpenAI formalized its go-to-market structure by launching the OpenAI Partner Network, a program backed by $150M to help partners build, sell, and deploy AI solutions. On the product side, ChatGPT added rate-limit reset banking for Plus and Pro users, and a Developer mode for the in-app browser that exposes controlled Chrome DevTools Protocol access for debugging. The ChatGPT memory management page now lets users delete individual memories or turn memory off entirely.