Claude Fable 5.2 is Anthropic's new model for demanding reasoning and long-horizon agent work, priced below Claude Fable 5.1. Changing the model ID is not the whole job. Four API changes can turn a working integration into 400 errors, so check how your code edits history, routes between models, calls computer use and pairs advisors first.
What was released
Claude Fable 5.2 (claude-fable-5-2) was announced by Anthropic on 9 November 2026 in the Claude API release notes. A sibling model, Claude Mythos 5.2 (claude-mythos-5-2), has the same capabilities but is open only to organisations verified through Anthropic's programmes, such as the Cyber Verification Program.
According to What's new in Claude Fable 5.2, both models have a 1M token context window, 128k maximum output tokens and adaptive thinking that is always on, with the effort parameter controlling depth. All five effort levels are supported and the API default is high. Fable 5.2 is available to all Claude API customers, and on Claude Platform on AWS, Amazon Bedrock (anthropic.claude-fable-5-2), both Google Cloud options and Microsoft Foundry. GitHub also added Claude Fable 5.2 to Copilot for Pro+, Max, Business and Enterprise plans, with an admin policy that is off by default.
What changed on price
Anthropic lists lower input and output prices than Fable 5.1, with the same multiplier for cache reads. Batch processing is half the standard rate. Fast mode is available on the Claude API with speed: "fast" and no beta header, at double the standard price for higher output speed.
Anthropic's own advice is worth noting: for most workloads, start with the latest Claude Opus model, and move to Fable 5.2 when your evals on Opus at higher effort still fall short.
The four breaking changes
1. The prefix check now applies to every account. If your code changes the system prompt, the tools or an earlier message, then sends back a later Fable 5.2 thinking block, the request returns a 400 error saying The block is bound to a different conversation. On Fable 5.1 this only applied to accounts created on or after 31 August 2026, so code that runs cleanly on an older account can start failing after the switch. A beta header, thinking-binding-controls-2026-08-01, with prefix_mismatch_behavior set to "drop_block" drops the affected blocks instead, but Anthropic calls that a stopgap because it can raise output tokens.
2. Only Fable 5.2 and Mythos 5.2 read Fable 5.2 thinking blocks. Moving a conversation onto Fable 5.2 keeps earlier reasoning from Opus 5.5, Sonnet 5.5, Haiku 5.5 and others. Moving it off Fable 5.2 to any other model, including Opus 5.5, Fable 5.1 and the fallback models Opus 4.8 and Opus 5, silently drops that reasoning. The request still succeeds.
3. Computer use must use the toolset. Fable 5.2 accepts computer_toolset_20260801. Declaring the earlier computer_20251124 tool returns a 400 error on every platform except Amazon Bedrock.
4. Advisor pairs are stricter. With the advisor tool (beta), a Fable 5.2 executor accepts only a Fable 5.2 or Mythos 5.2 advisor. Any other advisor, Fable 5.1 included, returns a 400 error. Changing the advisor's model counts as an edit to tools, so make the change in new conversations.
Two older rules also apply to every request: thinking set to enabled with a budget, or to disabled, returns a 400 error, and forced tool use through tool_choice is not supported. Coming from Fable 5.1 or Opus 5.5 these are not new, but coming from Opus 5 or earlier they are.
What it means for teams building agents
Audit how your code builds requests. The prefix check is the change most likely to surprise a team. Nothing changes if Claude Code, the Claude Agent SDK or Claude Managed Agents builds your requests. If your own code builds them, look for anything that trims old messages, rewrites the system prompt per turn, swaps tools mid-run or redacts earlier content. The migration guide has a test for this, and the docs point to compaction on demand and mid-conversation tool changes (both beta) as ways to keep history append-only. Our guide to context engineering for AI agents covers trimming strategies that now need a second look.
Review routers and fallbacks. If a router sends easy turns to a cheaper model, or a fallback catches refusals, every hop off Fable 5.2 loses its reasoning for those turns. That is a quality risk, not an error, so it only shows up in evals. Fable 5.2 runs safety classifiers, and refused requests return stop_reason: "refusal"; the permitted fallback targets are Opus 4.8 and Opus 5.
Update computer use and advisors in the same release. Teams that adopted the computer use toolset, as covered in Claude computer use toolsets, are ready. Anyone still on computer_20251124 outside Bedrock is not. Set the advisor model in the same change as the executor model, or the first request fails.
Check data retention before procurement does. Fable 5.2 requires 30-day data retention and is not available under zero data retention unless Anthropic expressly authorises it. Thinking blocks also work only in the account that produced them or a linked one. If your contracts promise zero retention, this decides the question before any eval does.
Re-run evals with an effort sweep. Anthropic recommends starting at high and moving to xhigh only for hard, long-running work where you have measured a gain. Prompt lines asking for less thinking don't stop Fable 5.2 from thinking at medium and above, and turns can end early on long tasks at lower effort. Watch subagent use too, since it can drive session cost. Our notes on what AI agent evals catch and controlling AI agent running costs apply here.
When not to switch. Stay on Opus if it already passes your evals, since Anthropic itself points most workloads there. Wait if you depend on zero data retention, if a cross-model router is central to your design, or if your agent edits history and you have not yet fixed it. If you are moving off Fable 5.1, plan it as you would any model deprecation: a staged rollout with evals at each step.
Frequently asked questions
Is Claude Fable 5.2 cheaper than Claude Fable 5.1?
Yes, per token. Anthropic lists lower input and output prices than Fable 5.1, with the same cache read multiplier and a half-price batch rate. Whether your bill falls depends on effort level and thinking volume, because the model thinks before almost every reply at high effort and above. Run your own workload at the default effort and compare total tokens per task, not only the price per token.
Why does my Claude Fable 5.2 request return a 400 prefix error?
Your code changed something before a thinking block it sent back: the system prompt, the tools or an earlier message. Fable 5.2 enforces this check on every account, unlike Fable 5.1. Fix the code so history is append-only and each assistant turn is returned exactly as received. As a short-term measure, the thinking-binding-controls beta header with drop_block set lets the request continue without those blocks.
Can I route between Claude Fable 5.2 and Claude Opus 5.5 in one conversation?
You can, and requests still succeed, but reasoning only travels one way. Fable 5.2 reads thinking blocks from Opus 5.5 and other recent models. Opus 5.5 cannot read Fable 5.2 thinking blocks, so the API drops them on any turn routed there. If your router moves turns back and forth often, measure answer quality on those turns before relying on it.
Can Claude Fable 5.2 be used under zero data retention?
Not by default. Anthropic states that Fable 5.2 and Mythos 5.2 require a retention period and are not available under zero data retention unless Anthropic expressly authorises it. In GitHub Copilot, business administrators must turn on a policy that acknowledges retention. Teams with zero retention commitments should check their agreements with their account team before planning a migration.
