Claude Fable 5 vs Fable 5.1: What Changed and How to Migrate

Anthropic released Claude Fable 5.1 on September 1, 2026. New integrations should use "claude-fable-5-1" rather than starting on the older "claude-fable-5" model ID.

This page now focuses on migration from Fable 5. For current setup instructions, prices, code samples, and supported platforms, use the complete Fable 5.1 API guide.

Fable 5.1 keeps the 1-million-token context window, 128,000-token maximum output, and standard $10 input / $50 output price per million tokens. It also uses the same tokenizer as Fable 5.

Your Anthropic API key, Messages API endpoint, SDK, and basic user-message structure remain the same. In a simple integration without custom thinking or forced tool settings, the model-name change may be the only code edit required.

Fable 5.1 always uses adaptive thinking. Requests that attempt to disable it or supply an unsupported manual thinking budget return 400. Assistant-message prefilling is also unsupported.

Forced tool choice is no longer accepted. If tool_choice is "any" or names a specific tool, the API returns 400. Use "auto" or "none" and enforce tool argument shape with strict tool use or structured outputs.

Thinking blocks require careful history handling. Earlier Claude models cannot read Fable 5.1 thinking blocks, and editing a previous turn, system message, or tool definition may invalidate them. Keep conversation history append-only and preserve returned blocks exactly.

Standard input and output prices remain $10 and $50 per million tokens. The important cost change is the cache-read rate: Fable 5.1 charges $0.25 per million cached input tokens. A 5-minute cache write costs $12.50 and a 1-hour write costs $20.

Anthropic estimates that typical workloads can cost about 25% less than Fable 5 after accounting for cache reads, while highly agentic workloads can save up to roughly 45%. Those are vendor estimates. Measure your own cache-hit rate and completed-task cost before using them in a budget.

Start with a copy of production request logs that has API keys and private content removed. Identify every request that sets thinking, tool_choice, assistant prefills, or replays edited history. Update those paths before changing the default model.

Run Fable 5 and Fable 5.1 against a fixed evaluation set. Compare completed-task quality, latency, token use, cache-hit rate, tool errors, refusals, and total cost. A model can have a higher per-token price and still cost less per completed task, but that must be measured.

Roll out to a small percentage of traffic, monitor 400 errors and fallback behavior, then expand. Keep an explicit rollback path that starts a clean conversation with the older model instead of passing incompatible Fable 5.1 thinking blocks backward.