Claude Fable 5.2: pricing, thinking and migration changes

Claude Fable 5.2 is Anthropic's newest general-purpose model in the Fable line, released on 9 November 2026 under the API ID claude-fable-5-2. It is cheaper than Fable 5.1 at $8 per million input tokens and $40 per million output tokens, keeps the 1M-token context window and 128k output limit, and runs adaptive thinking on every request. It also ships four API changes that can break code written for Fable 5.1, so a model-string swap is not the whole migration.

This post covers what the model is, what it costs, how thinking and effort behave, and the specific request shapes that now return errors.

What is Claude Fable 5.2?

Fable is Anthropic's top tier for long, hard work: agentic coding, multi-step research, computer use and long-running automation. Version 5.2 replaces Fable 5.1 (claude-fable-5-1) and is available to all Claude API customers. Anthropic describes it as a general capability step up, with outputs up to 30% faster than Fable 5.1 and lower token use at medium effort.

Alongside it, Anthropic released Claude Mythos 5.2 (claude-mythos-5-2). It is the same underlying model with some domain safeguards removed, offered only to organizations verified through Anthropic's access programs. For most developers, Fable 5.2 is the one you will call.

The core specs:

• Context window: 1M tokens, which is both the default and the maximum.
• Max output: 128k tokens per response.
• Knowledge cutoff: June 2026 (reliable cutoff, per the models overview).
• Platforms: the Claude API, Amazon Bedrock (bedrock-runtime only), Google Cloud, and Microsoft Foundry on Anthropic infrastructure.
• Data retention: Fable 5.2 requires 30-day retention and is not available under zero data retention unless Anthropic authorizes it.

How much does Claude Fable 5.2 cost?

Per million tokens:

That is a 20% cut on both input and output compared with Fable 5.1. Cache reads stay at 0.025x the base input price, the same multiplier as before. Batch is half price, and fast mode, which Anthropic says produces output about 2.5x faster, costs double and cannot be combined with the Batch API.

One thing that does not change: Fable 5.2 uses the same tokenizer as Claude Opus 4.7 and later. A tokenizer splits text into the units you are billed for, and this one produces roughly 30% more tokens for the same text than older Claude models. If you are coming from Fable 5.1 your counts carry over. If you are coming from an older model, recount with real prompts. The Haiku 5.5 pricing breakdown walks through that arithmetic.

How do thinking and effort work on Fable 5.2?

Adaptive thinking means the model decides how much to reason before answering. On Fable 5.2 it is always on. You cannot turn it off, and you cannot give it a fixed budget:

• thinking: {"type": "enabled", "budget_tokens": N} returns a 400.
• thinking: {"type": "disabled"} also returns a 400.
• Omit thinking, or send {"type": "adaptive"}.

Effort is the lever you do get. It tells the model how many tokens to spend across thinking, text and tool calls. There are five levels: low, medium, high, xhigh and max. The API default is high. Note that Claude's own apps and Claude Code default to medium, so results you see there may not match your API calls at default settings.

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-fable-5-2",
    max_tokens=32000,
    output_config={"effort": "high"},
    messages=[{"role": "user", "content": "Refactor this module and explain the changes: ..."}],
)

text = "".join(b.text for b in response.content if b.type == "text")
print(text)

Anthropic's guidance is to start at high, move to xhigh for the hardest long-running jobs, and keep max for finding the ceiling in evals. Two behaviours are worth knowing. Asking for less thinking in the prompt does not stop thinking at medium and above; change effort instead. And at low and medium, a turn is more likely to end before the work is finished, so agent loops should check for completion rather than assume it.

What breaks when you migrate from Fable 5.1?

The docs list four changes that can break existing code.

1. Thinking blocks are checked against the conversation. A thinking block is the reasoning content the model returns, which you send back in later turns. On Fable 5.2 every account gets a prefix check: if a request contains a thinking block that does not match the conversation it came from, you get a 400 whose message includes "The block is bound to a different conversation". This catches code that edits, reorders or splices history. The stopgap is the thinking-binding-controls-2026-08-01 beta header with prefix_mismatch_behavior set to "drop_block":

{
  "model": "claude-fable-5-2",
  "max_tokens": 32000,
  "thinking": {
    "type": "adaptive",
    "block_binding": { "prefix_mismatch_behavior": "drop_block" }
  },
  "messages": [ ... ]
}

The response then lists anything it dropped in input_transformations. Dropping blocks can raise output tokens, because the model has to redo that reasoning, so the real fix is to stop editing history that contains thinking.

2. Thinking blocks are bound to the model. Only Fable 5.2 and Mythos 5.2 can read Fable 5.2 thinking blocks on the Claude API and Google Cloud's Google-managed endpoint. If you switch mid-conversation to another model, such as Opus 5.5 or Fable 5.1, those blocks are dropped. You are not billed for dropped blocks, but the new model loses that context.

3. The old computer use tool is rejected. Requests that declare computer_20251124 return a 400 on every platform except Amazon Bedrock. Use computer_toolset_20260801 instead.

4. Advisors must match. If you use a Fable 5.2 executor with an advisor tool, the advisor must be claude-fable-5-2 or claude-mythos-5-2. Fable 5.1 or Mythos 5.1 advisors return a 400. Changing the advisor is a change to tools, so it needs a fresh conversation.

Two more restrictions are easy to miss:

Refusals arrive as HTTP 200 with stop_reason: "refusal" and a stop_details object. A beta fallbacks: "default" option can retry on a permitted fallback model server-side; otherwise, check stop_reason in your client.

When should you use Claude Fable 5.2?

It fits work where quality on long, multi-step tasks matters more than cost per call: agentic coding across a large repository, long research runs, computer use, and jobs that run for hours. The 1M context and 128k output make it practical to keep a whole codebase or document set in one conversation.

For narrow, high-volume tasks such as classification, routing or extraction, a smaller model is still the cheaper fit. A common split is Fable 5.2 for planning and final output, with a smaller model handling well-scoped subtasks.

If you are already on Fable 5.1, the price cut makes the move attractive, but treat it as a migration. Change the model string, search your code for history edits, model switches, computer_20251124, advisor models and forced tool_choice, then rerun your evals with an effort sweep starting at high.

Key takeaways

• Claude Fable 5.2 (claude-fable-5-2) costs $8/$40 per million tokens, 20% less than Fable 5.1.
• 1M context, 128k output, and adaptive thinking that cannot be disabled or budgeted.
• Effort defaults to high on the API; tune it rather than prompting for less thinking.
• Editing history with thinking blocks, forced tool_choice and computer_20251124 all return 400s.
• Thinking blocks do not carry over if you switch to a different model mid-conversation.

FAQ

Is Claude Fable 5.2 a drop-in replacement for Fable 5.1?

Not quite. The model string change is simple, but four API changes can return errors, mainly around thinking blocks, the computer use tool version, advisor models and forced tool choice.

Can I turn off thinking on Claude Fable 5.2?

No. Both enabled with a budget and disabled return a 400. Lower the effort level to reduce how much the model thinks.

What is the difference between Fable 5.2 and Mythos 5.2?

They share the same underlying model. Mythos 5.2 has some domain safeguards removed and is available only to organizations verified through Anthropic's access programs.