Claude Fable 5.2 for Builders: What Changed, and What to Do About It

Anthropic released Claude Fable 5.2 on 9 November 2026, alongside Claude Mythos 5.2, its twin for verified organizations. For people who use the API, the headline is easy: it is the most capable model Anthropic offers to everyone, and it is cheaper than the model it replaces, $8 per million input tokens and $40 per million output, down from $10 and $50 for Fable 5.1.

The details are where projects break. Thinking is always on, effort is the only control, and a check that used to apply only to newer accounts now applies to every account. This guide condenses Anthropic's Fable 5.2 documentation into what API builders need to change, with the parts most likely to break production code first.

The model at a glance

	Claude Fable 5.2
Model id	claude-fable-5-2 (no date suffix)
Price	$8 input, $40 output per million tokens; batch $4 and $20
Cache	Reads $0.20 per million; 5-minute writes $10; 1-hour writes $16
Fast mode	$16 and $80, on the Messages API (not in Managed Agents)
Context	1M tokens, which is both the default and the maximum
Max output	128K tokens
Thinking	Always on; control depth with output_config.effort, default high
Data retention	30-day retention required; not available under zero data retention unless Anthropic authorizes it

Anthropic also reports output up to 30% faster than Fable 5.1. Claude Opus 5.5, at $4 and $20, stays the sensible default for most workloads. Fable 5.2 is for the work where Opus at high effort still falls short.

Effort is the only dial

On Fable 5.2 you cannot turn thinking off, and you cannot give it a thinking budget. thinking: {type: "disabled"} and budget_tokens both return a 400. Sampling parameters such as temperature are gone too. What you control is effort: five levels from low to max, with high as the default.

Turn the dial to see what Anthropic recommends each level for, and the request that sets it:

const stream = client.beta.messages.stream({
  model: 'claude-fable-5-2',
  max_tokens: 64000,
  output_config: { effort: 'high' },
  betas: ['server-side-fallback-2026-07-01'],
  fallbacks: 'default',
  messages,
});
const msg = await stream.finalMessage();
if (msg.stop_reason === 'refusal') { /* handle it */ }

Three behaviors are worth knowing before you pick:

• Instructions to think less no longer work at medium and above. Lines like "answer directly without deliberating" used to trim thinking on Fable 5.1. Delete them and lower the effort instead.
• Thinking counts toward max_tokens. A request sized tightly for a short answer can now stop with stop_reason: "max_tokens" after a thinking block and before any text. Give high-effort requests room and stream them.
• Effort barely changes reply length. If you want shorter answers, say so in the prompt. Effort controls how hard the model works, not how much it writes.

Anthropic's advice is to start at high, then run your evals at medium and xhigh too. Most of the capability gain arrives by high; gains above it are uneven.

The change most likely to break you: history is checked

Every thinking block Fable 5.2 produces is bound to the system prompt, tools and messages that came before it. Send that block back after editing any of them, and the API returns a 400 saying the block is bound to a different conversation. Retrying the same body does not help.

Fable 5.1 and Opus 5.5 already did this, but only for accounts created on or after 31 August 2026. On Fable 5.2, every account is checked. If your account is older, code that works on Fable 5.1 today can fail on Fable 5.2 tomorrow.

You are fine if Claude Code, the Claude Agent SDK, Managed Agents or claude.ai manages the conversation, or if your code keeps system and tools fixed for a session and only appends to messages. You need work if your harness:

• rewrites the system prompt mid-session, for example to inject the time or a user's state;
• adds or removes tools between turns, including an MCP server whose tool list changes;
• trims or summarizes old messages by editing them in place.

The fix is to keep history append-only. Freeze system and tools per session, put instruction changes in mid-conversation system messages, and trim with server-side compaction or context editing. Ship prompt changes in new conversations.

A related change: no model released before Fable 5.2 can read its thinking blocks. If a router or a refusal fallback moves a conversation to Opus 5.5 or Fable 5.1, the turns after the switch run without that reasoning. The request succeeds; it just costs more on the first turn after the switch.

Your migration, filtered

What you need to change depends on where you are coming from. Pick your current model and tick items off as you go:

Handle refusals in code

Fable 5.2 runs safety classifiers that can decline a request with HTTP 200 and stop_reason: "refusal", plus a stop_details object with a category such as cyber, bio or reasoning_extraction. Always check stop_reason before you read content.

You can also have the API retry a declined request on another model for you. With the beta header server-side-fallback-2026-07-01, set fallbacks: "default" and the API routes each category to the model Anthropic recommends for it. Remember the point above: the fallback model does not see Fable 5.2's thinking.

One easily missed category is reasoning_extraction. If a prompt or tool description asks the model to write out its reasoning in the reply, remove that line. Read display: "summarized" thinking blocks instead.

Where Fable 5.2 belongs in a system

At $8 and $40 per million tokens, plus thinking tokens billed as output, Fable 5.2 is the most expensive way to make a simple judgment. Asking it whether an email is spam works, but it costs thousands of times what a decision model charges for the same answer.

Our rule has not changed with the new model: a decision model decides, a frontier model writes. Jev reads every message and answers the typed questions; Fable 5.2 only sees the messages that need careful writing, reasoning or code. We show that architecture running, with the costs, in Jev decides, Fable 5.2 writes.

If you are planning a Fable 5.2 migration or a multi-model system, our AI engineering services cover exactly this, or start a conversation.

FAQ

What is the Claude Fable 5.2 model id?

claude-fable-5-2, with no date suffix. On Amazon Bedrock, where offered, it is anthropic.claude-fable-5-2.

How much does Claude Fable 5.2 cost?

$8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1. Cache reads cost $0.20 per million tokens, batch processing is half price, and fast mode costs $16 and $80.

Can I turn off thinking on Claude Fable 5.2?

No. Thinking is always on, and thinking: {type: "disabled"} returns a 400, as does budget_tokens. Lower output_config.effort to low or medium to reduce thinking.

Why does Fable 5.2 return a 400 about a thinking block bound to a different conversation?

Your code edited the system prompt, tools or an earlier message and then sent back a thinking block produced before the edit. Fable 5.2 enforces this check on every account. Keep history append-only and make prompt changes in new conversations.