Impact note · revision 1.0
Fable 5.1 rejects forced tool choice
Requests using tool_choice type any or tool now fail before inference; auto and none remain supported.
Bottom line
A request with tool_choice set to any or a named tool returns a 400 invalid_request_error. Retrying the same body does not make it compatible.
Operator action
What to do
Replace forced tool choice with auto plus an explicit instruction and a strict tool schema, or use structured outputs when the real requirement is schema-valid JSON.
- When
- Before sending production traffic to claude-fable-5-1
- Evidence
- Officially confirmed
- Last verified
Who is affected
- Claude Messages API
- Message Batches API
- Token counting endpoint
- claude-fable-5-1
- claude-mythos-5-1
What changed
- Claude Fable 5 accepts auto, none, any, and named-tool choices; Fable 5.1 accepts only auto and none.
- The same validation applies to token counting as well as message execution.
- Anthropic recommends auto with strict tool use, or structured outputs for a fixed response shape.
Migration checks
- Search request builders for tool_choice values any and tool.
- Move the requirement into the user turn or an appended system message.
- Set strict: true on the tool schema when that feature is available.
- Exercise the no-tool path as well as the expected tool path before cutover.
What we observed
The bundled Node.js preflight was executed against six frozen synthetic fixtures. It checks only the encoded request-shape and prefix rules. No Anthropic API request was made, and no production behavior was reproduced.
Inspect the local preflight →What remains unknown
- This pilot has not measured tool-call reliability for any production prompt set.
- CMEK environments have documented structured-output constraints that require separate verification.
Primary evidence
- Claude Fable 5.1 overviewAnthropic primary source · checked Sep 11, 2026
- Migrating to Claude Fable 5.1 and Claude Mythos 5.1Anthropic primary source · checked Sep 11, 2026
- Claude API errorsAnthropic primary source · checked Sep 11, 2026
Revision and corrections
- Initial pilot edition; no material correction recorded.
Material corrections are dated and described; records are not silently rewritten. Read the correction policy.