Day 0 Support: Claude Fable 5.1

LiteLLM supports Claude Fable 5.1 on Day 0, with spend, rate limits, fallbacks, and logging in one place. Anthropic is live now; Bedrock, Vertex AI, and Azure AI land later today.

What's new in Fable 5.1

Pricing, the 1M-token context window, the 128K output ceiling, and always-on adaptive thinking carry over from Fable 5. What changed, in the order it costs a gateway work:

• Cache reads cost $0.25 / MTok, down from $1.00. That is 0.025x base input, where every other Claude model sits at 0.1x. LiteLLM reads the rate from the cost map and falls back to base input without it, a 40x overcount.
• Forced tool use returns a 400. Thinking is always on, and a forced call would skip it. LiteLLM maps OpenAI's tool_choice: "required" to Anthropic's any, so a request that worked on Fable 5 fails here unchanged. Keep auto with strict: true for schema-valid JSON, or name the tool in the prompt to force a call.
• Thinking blocks are bound to the model that wrote them. Fable 5.1 reads earlier models' blocks; none of them read its own.
• Editing earlier turns invalidates thinking blocks. Rebuilding system or tools mid-conversation errors with The block is bound to a different conversation, on accounts created from August 31, 2026.
• Effort is steerable per message behind the mid-conversation-output-config-2026-07-01 beta header, without invalidating the prompt cache.

Anthropic's what's new page has the benchmarks and capability gains.

Usage

Anthropic

model_list:
  - model_name: claude-fable-5-1
    litellm_params:
      model: anthropic/claude-fable-5-1
      api_key: os.environ/ANTHROPIC_API_KEY

Azure

model_list:
  - model_name: claude-fable-5-1
    litellm_params:
      model: azure_ai/claude-fable-5-1
      api_key: os.environ/AZURE_AI_API_KEY
      api_base: os.environ/AZURE_AI_API_BASE  # https://<resource>.services.ai.azure.com

Vertex AI

model_list:
  - model_name: claude-fable-5-1
    litellm_params:
      model: vertex_ai/claude-fable-5-1
      vertex_project: os.environ/VERTEX_PROJECT
      vertex_location: global

Pinned regions carry a 10% premium over global, which LiteLLM applies.

Bedrock

model_list:
  - model_name: claude-fable-5-1
    litellm_params:
      model: bedrock/converse/us.anthropic.claude-fable-5-1
      aws_access_key_id: os.environ/AWS_ACCESS_KEY_ID
      aws_secret_access_key: os.environ/AWS_SECRET_ACCESS_KEY
      aws_region_name: us-east-1

Bedrock serves it through inference profiles. us. and eu. carry the 10% regional premium, global. stays at base.

curl --location 'http://0.0.0.0:4000/chat/completions' \
--header 'Content-Type: application/json' \
--header 'Authorization: Bearer $LITELLM_KEY' \
--data '{
  "model": "claude-fable-5-1",
  "messages": [{"role": "user", "content": "what llm are you"}],
  "reasoning_effort": "xhigh"
}'

reasoning_effort maps to adaptive thinking, the only mode Fable 5.1 accepts. Fixed budgets, assistant prefill, and non-default temperature or top_p return a 400. Pass output_config: {"effort": "max"} for the full ladder Fable 5 uses.

Before you switch traffic over

Thinking blocks travel one way, so a fallback onto Fable 5.1 keeps the conversation's reasoning while one off it to Opus 5 or Opus 4.8 drops it for the turns that run there. Keep the chain regardless: refusals come back as HTTP 200 with stop_reason: "refusal", and the permitted targets are Opus 4.8 and Opus 5. Treat history as append-only, since editing an earlier turn invalidates every thinking block after it.

Four behaviors also differ with no code change on your side: one tool call per turn where Fable 5 batched several, fewer progress updates between tool calls, more answers from memory at low effort, and whole-file rewrites in place of targeted edits. Anthropic's prompting guide has a fix for each.

Feedback

Hitting something unexpected? Share it on GitHub discussion #39163.