Claude Fable 5.1: Three Breaking API Changes to Fix Now

Anthropic shipped Claude Fable 5.1 on September 1. Most developers updated the model ID string, ran a smoke test, and moved on. That was the right call for simple prompt-response integrations. For anyone running tool-calling agents, multi-turn conversations, or multi-model routers, it was not enough — and some of those codebases are now throwing 400 errors in production.

There are three breaking API changes in Fable 5.1. Each one is fixable in an afternoon. None of them are optional.

The Good News First: Cache Reads Are 75% Cheaper

Cache read pricing dropped from $1.00 to $0.25 per million tokens. Input and output prices held at $10 and $50 per million respectively. The savings are automatic — no code change required. For agents that repeatedly send the same system prompt, codebase context, or documentation, Anthropic estimates 25% savings on typical workloads and up to 45% on heavily agentic ones. That is a meaningful reduction at scale.

Breaking Change 1: Forced Tool Use Now Returns 400

If your code sends tool_choice: {type: "tool"} or tool_choice: {type: "any"}, Fable 5.1 rejects it immediately with a 400 error on the Messages API, Batches API, and the token-counting endpoint. This is not a deprecation warning — it breaks on the first request.

The fix: switch to tool_choice: {type: "auto"} and move your tool selection intent into the system prompt.

# Breaks on Fable 5.1
response = client.messages.create(
    model="claude-fable-5-1",
    tool_choice={"type": "tool", "name": "search"},
    tools=[search_tool],
    messages=messages
)

# Correct
response = client.messages.create(
    model="claude-fable-5-1",
    tool_choice={"type": "auto"},
    system="Always use the search tool for factual lookups.",
    tools=[search_tool],
    messages=messages
)

The model picks the right tool when you tell it to. Forced selection was always a workaround; Anthropic is closing it.

Breaking Change 2: Conversation History Must Be Append-Only

This is the change most likely to catch framework users off guard. If a conversation contains a thinking block and you later edit, reorder, or remove any earlier turn, every thinking block after that point is invalidated. This is Anthropic’s anti-distillation measure — it closes the documented technique of manipulating context around thinking blocks to extract model reasoning.

Enforcement is strict for API accounts created on or after August 31, 2026. Older accounts receive logged warnings for now, but enforcement is coming.

According to Digital Applied’s framework compatibility analysis, six of ten major agent frameworks break here by default: LangChain, LangGraph, Vercel AI SDK, OpenClaw, Mastra, and Semantic Kernel all do history trimming or summarization that keeps a verbatim tail while prepending a summary. That pattern invalidates thinking blocks in the preserved tail.

The safe approaches: treat conversations as strictly append-only, or if you must compress history, replace the entire history with a single summary block. Partial replacement with a preserved tail is what breaks.

Breaking Change 3: Thinking Blocks Do Not Cross Model Boundaries

Claude Fable 5.1 reads thinking blocks produced by earlier models without issue. The reverse is not true: Fable 5 and Opus 5 cannot read Fable 5.1 thinking blocks. If you have a multi-model router with a Fable 5 fallback, or an A/B test running both versions concurrently, any Fable 5.1 thinking blocks passed to an earlier model will be dropped silently or cause an error.

The fix: strip thinking blocks before routing to an earlier model, or set the drop_block flag in your configuration. If your fallback logic exists for reliability, make sure it strips state the downstream model cannot process.

Who Does Not Need to Change Code

Claude Managed Agents users: update the model name only. Stateless endpoints with no tools and no stored conversation history are also safe with a model ID update alone. Simple, single-turn prompt-response chains are unaffected by all three changes.

If you use the Anthropic Python or TypeScript SDK without tool_choice overrides and without editing conversation history, you most likely only need to update the model name and run your regression suite.

How to Migrate

1. Search your codebase for tool_choice. Replace "type": "tool" and "type": "any" with "type": "auto" and add tool-use intent to the system prompt.
2. Audit conversation history handling. Make it append-only or use full-replacement summarization — not tail-preserve.
3. Audit multi-model routing. Strip Fable 5.1 thinking blocks before sending to older models.
4. Run offline regression tests, then shadow traffic, then a limited canary. Keep a config-level rollback to the previous model ready until you are confident.

Anthropic’s official Fable 5.1 migration guide covers all three changes with complete examples. The cache price cut is automatic — you get that by just updating the model name. The three breaking changes are not.

The developers who treated this as a model ID swap will find out in production. Do the migration.