Claude Fable 5.1 and Mythos 5.1 land; tool_choice 'any' now 400s
Anthropic's agentic Mythos-class models get a point upgrade: prompt-cache reads drop to $0.25/MTok, thinking is always on, and old tool_choice calls break.
Copy markdown
Two new IDs, one you probably can't touch
`claude-fable-5-1` is live on the public API; `claude-mythos-5-1` stays gated to Project Glasswing (government cyberdefenders and selected bio researchers). Both are Mythos-class agentic models tuned for long-horizon coding — Fable 5 already topped Cognition's FrontierCode among frontier models and reportedly lands SWE-bench Verified in the mid-90s.
Cache reads just got 4x cheaper
Prompt-cache reads drop to $0.25/MTok — 0.025x the base input price, versus the usual 0.1x. If your agent replays a fat system prompt or a million-token context every turn, that's the line item that actually moves your bill. Base rates hold at $10 in / $50 out per MTok.
1M context, and thinking you can't switch off
Both models default to a 1M-token window with up to 128k output tokens, and adaptive thinking is always on — there's no non-thinking mode to drop into. Budget for the extra output tokens and the added latency inside tight agent loops.
tool_choice 'any' and 'tool' now return 400
Force a tool call with `tool_choice: {type:"any"}` or `"tool"` and 5.1 rejects it — only `auto` and `none` are accepted. Move to strict tool use or structured outputs when you need to guarantee a call, and audit your agent scaffolds before you swap the model ID in.
Replayed thinking blocks can 400 mid-run
5.1 reads thinking blocks from Opus 5, Fable 5, Mythos 5 and earlier, but for accounts created on or after Aug 31 it errors if you replay a block after the system prompt, tools, or a prior message changed. Multi-turn agents that mutate context need to stop echoing stale thinking.
Where you can call it, and the retention catch
Available now on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. The catch: Fable 5.1 requires 30-day data retention and won't run under zero-data-retention without Anthropic's explicit sign-off.
Elsewhere: OpenAI starts sunsetting Whisper
OpenAI deprecated `whisper-1` and the gpt-4o-transcribe family on Aug 26, with a hard shutdown on Feb 26, 2027; new builds should target `gpt-live-transcribe` or `gpt-transcribe`. If you've got a Whisper pipeline in production, schedule the swap now.