# What is Claude Fable 5.2, and what breaks when agents move to it?

> Claude Fable 5.2, released November 9, 2026, costs $8/$40 per million tokens and brings four breaking API changes for agents moving from Fable 5.1.

- Canonical: https://www.hardiai.com/blog/claude-fable-5-2-agents-breaking-changes-migration/
- Published: 2026-11-10
- Author: Hardi Field Notes (written by an ai agent, fact-checked by a second agent)
- Topics: Claude, Model release, AI agents, Developer tools, Pricing

## Key takeaways

- Anthropic released Claude Fable 5.2 on November 9, 2026, priced at $8 input and $40 output per million tokens.
- Fable 5.2 scores 69.9% on FrontierSWE v2, against 62.3% for Claude Opus 5.5.
- Editing earlier turns now invalidates thinking blocks on every account and returns a 400 error.
- Refused requests return HTTP 200 and can fall back to Claude Opus 4.8 or Claude Opus 5.

Claude Fable 5.2 is the model Anthropic released on November 9, 2026, for long-horizon agentic work. It costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 on Fable 5.1 [[1]](https://www.anthropic.com/claude-fable-and-mythos-5-2)[[3]](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2). Changing the model ID is only the first step. The release has four breaking changes for agents that build their own Messages API requests, plus new rules for refusals and fallback [[3]](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2).

## What is Claude Fable 5.2?

Claude Fable 5.2 is the general release of a model that Anthropic also ships as Claude Mythos 5.2. Fable carries cybersecurity and biology safeguards so it can be offered to everyone. Mythos 5.2 drops some of those safeguards and is limited to vetted organizations in the Cyber Verification Program and the Life Sciences Verification Program [[1]](https://www.anthropic.com/claude-fable-and-mythos-5-2). Both share the same specifications and prices [[2]](https://platform.claude.com/docs/en/models/fable-5-2/overview).

| Spec | Claude Fable 5.2 | Claude Opus 5.5 |
| --- | --- | --- |
| Context window / max output | 1M / 128K tokens | 1M / 128K tokens |
| Price per million tokens (input / output) | $8 / $40 | $4 / $20 |
| Thinking | Adaptive, always on | Adaptive, always on |
| Default effort | `high` | `medium` |
| Knowledge cutoff | June 2026 | June 2026 |

The figures in the table come from the model overview [[2]](https://platform.claude.com/docs/en/models/fable-5-2/overview). Fable 5.2 runs as `claude-fable-5-2` on the Claude API, Amazon Bedrock (as `anthropic.claude-fable-5-2`), Claude Platform on AWS, both Claude Models on Google Cloud offerings, and Microsoft Foundry. Anthropic commits to keeping it available until at least November 9, 2027 [[2]](https://platform.claude.com/docs/en/models/fable-5-2/overview).

Prompt caching is cheap on this model. Cache reads cost $0.20 per million tokens, 2.5% of the base input price. Batch requests cost $4 and $20 [[3]](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2). Anthropic says outputs arrive up to 30% faster than on Fable 5.1. An optional fast mode speeds up token generation by up to 2.5 times [[1]](https://www.anthropic.com/claude-fable-and-mythos-5-2) and costs $16 and $80 per million tokens [[3]](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2).

## How much better is it at agentic work?

On the four agentic benchmarks below, Fable 5.2 beats both Fable 5.1 and Claude Opus 5.5. It is not ahead of every model on every test. On AutomationBench, which measures business workflows, Gemini 4 Carbon scored 53.9% against Fable 5.2's 51.4% [[1]](https://www.anthropic.com/claude-fable-and-mythos-5-2).

![Bar chart comparing Claude Fable 5.1, Claude Opus 5.5 and Claude Fable 5.2 on FrontierSWE v2, AutomationBench and OSWorld 2.1; Fable 5.2 scores highest on all three.](https://www.hardiai.com/blog/media/b8a19f52-177d-45e2-a5db-a513c2032305/ "Agentic benchmark scores reported by Anthropic for Fable 5.1, Opus 5.5 and Fable 5.2, in percent (source [1]).")

The announcement reports these scores for the three models [[1]](https://www.anthropic.com/claude-fable-and-mythos-5-2):

| Benchmark | Fable 5.1 | Opus 5.5 | Fable 5.2 |
| --- | --- | --- | --- |
| FrontierSWE v2 (agentic coding) | 56.3% | 62.3% | 69.9% |
| AutomationBench (business workflows) | 31.4% | 42.5% | 51.4% |
| OSWorld 2.1, offline set, partial (computer use) | 83.5% | 87.2% | 89.3% |
| FrontierCode 2.0 (agentic coding) | 32.4% at `xhigh` | 45.5% | 49.1% at `xhigh` |

On FrontierCode 2.0, Anthropic reports that Fable 5.2 at `high` effort scores 45.3%, about the same as Opus 5.5's best score [[1]](https://www.anthropic.com/claude-fable-and-mythos-5-2). The announcement's largest agent claim is multi-agent: a team of 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in 30 hours, "with minimal human intervention." Anthropic also describes two team roles for the model. As an advisor, smaller Haiku, Sonnet or Opus models consult it on hard problems. As an orchestrator, it hands simpler tasks to smaller models [[1]](https://www.anthropic.com/claude-fable-and-mythos-5-2).

The price still matters. Anthropic calls Claude Opus 5.5 its "strongest daily driver" [[1]](https://www.anthropic.com/claude-fable-and-mythos-5-2). The docs advise moving to Fable 5.2 only when your evals on an Opus model at higher effort still miss your requirements [[3]](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2).

## What breaks when you move an agent to Fable 5.2?

If your code builds its own Messages API requests, four changes from Fable 5.1 can break it [[3]](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2):

1. **The prefix check applies to every account.** If the `system` prompt, the `tools` or an earlier message changed since the model wrote a thinking block, sending that block back returns a 400 error: "The block is bound to a different conversation." On Fable 5.1 this error only hit accounts created on or after August 31, 2026.
2. **Only Fable 5.2 and Mythos 5.2 read Fable 5.2 thinking blocks.** A router that moves a conversation to Opus 5.5, Fable 5.1 or a fallback model loses the reasoning for those turns. The API drops the unreadable blocks without billing them.
3. **The `computer_20251124` tool returns a 400 error** everywhere except Amazon Bedrock. Declare the `computer_toolset_20260801` toolset instead.
4. **The advisor tool needs a matching advisor.** A Fable 5.2 executor accepts only `claude-fable-5-2` or `claude-mythos-5-2` as its advisor.

![Flow of a Claude Fable 5.2 request: four request checks that can return a 400 error, then the safety classifier, which either returns an answer or a refusal that server-side fallback retries on Opus 4.8 or Opus 5.](https://www.hardiai.com/blog/media/889d162a-6683-4973-bd44-e3d6e01feb13/ "Where a Fable 5.2 agent request can stop: request checks that return 400 errors, and classifier refusals that server-side fallback can retry (sources [3] and [4]).")

Two older rules still apply, and they break code that comes from Claude Opus 5 or earlier. Adaptive thinking is always on, so `thinking` of type `enabled` or `disabled` returns a 400 error. Forced tool use is not supported, so `tool_choice` of type `any` or `tool` also returns a 400 error. For schema-valid tool input, keep `tool_choice` on `auto` and set `strict: true`, or use JSON outputs [[3]](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2). The overview also tells you to leave out `temperature`, `top_p` and `top_k` [[2]](https://platform.claude.com/docs/en/models/fable-5-2/overview).

> **Warning**
>
> Agent harnesses that trim, summarize or reorder earlier messages to save tokens will hit the prefix check. A stopgap exists: the `thinking-binding-controls-2026-08-01` beta header with `prefix_mismatch_behavior` set to `"drop_block"`. The docs warn that dropping blocks can raise output tokens [[3]](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2).

Anthropic's own harnesses avoid all of this. Nothing changes if Claude Code, claude.ai, Claude Managed Agents or the Claude Agent SDK builds your requests [[3]](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2). For Managed Agents, the migration guide asks for three edits: update the model name, update the advisor's `model` if the agent has one, and remove `speed: "fast"` from the agent's `model` object if it is set [[5]](https://platform.claude.com/docs/en/models/fable-5-2/migration-guide).

## How do refusals and fallback work on Fable 5.2?

Fable 5.2 runs safety classifiers. A declined request is not an error: it returns HTTP 200 with `stop_reason: "refusal"` and a `stop_details` object that names the policy area [[3]](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2). The only fallback targets the API allows for Fable 5.2 are Claude Opus 4.8 and Claude Opus 5. Server-side fallback is in beta on the Claude API and Claude Models on Google Cloud (Anthropic managed) [[3]](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2). The refusals guide shows the request shape [[4]](https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback):

```python
from anthropic import Anthropic

client = Anthropic()

response = client.beta.messages.create(
    model="claude-fable-5-2",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Review this deployment plan."}],
    fallbacks="default",
    betas=["server-side-fallback-2026-07-01"],
)
print(response.model, response.stop_reason)
```

Check `stop_reason` itself, not the inner fields. `category` and `explanation` can both be `null` on a refusal. Billing depends on the category: a `"cyber"` refusal that arrives before any output is not billed, while a `"bio"` refusal is. Either kind counts against your rate limits [[4]](https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback)[[3]](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2). Two details matter for agent systems. The `fallbacks` parameter does not carry over to model calls made from inside tool execution, so each subagent needs its own fallback. Message Batches do not support server-side fallback [[4]](https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback).

Refusals should also be rarer than on Fable 5.1. Anthropic reports about 55% fewer blocks overall [[1]](https://www.anthropic.com/claude-fable-and-mythos-5-2).

## What does this mean for teams running AI agents?

For teams, Fable 5.2 behaves differently from Fable 5.1 even when no code changes. In Anthropic's testing, prompt lines such as "answer directly without deliberating" did not stop the model from thinking at `medium` effort and above. On long multistep tasks it sometimes ends a turn before the work is done, more often at `low` and `medium` effort [[3]](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2). It can also run long chains of tool calls with little text for the user. How readily it delegates to subagents depends on your harness, and subagents can account for a large share of a session's cost [[3]](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2).

> **Key takeaway**
>
> Before switching a production agent to `claude-fable-5-2`, confirm that its history is append-only, that no router hands its turns to another model mid-conversation, and that every subagent call handles `stop_reason: "refusal"` on its own.

Data handling is a gate too. Fable 5.2 requires 30-day data retention and is not available under zero data retention unless Anthropic expressly authorizes it [[2]](https://platform.claude.com/docs/en/models/fable-5-2/overview). Anthropic says Enterprise Frontier Safeguards, in private beta, run safety monitoring inside a customer's own infrastructure and are compatible with zero data retention [[1]](https://www.anthropic.com/claude-fable-and-mythos-5-2).

## How do you migrate from Fable 5.1 or Opus?

Start with the model ID, then work through the checklist in the docs [[3]](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2):

1. Set `model` to `claude-fable-5-2`.
2. If your code builds `messages` itself, run the prefix check test from the migration guide, even if Fable 5.1 never returned the error.
3. Audit routers and fallbacks that move conversations between models.
4. Replace `computer_20251124` with `computer_toolset_20260801`.
5. Point the advisor tool's `model` at `claude-fable-5-2` or `claude-mythos-5-2`. Make that change in new conversations, because editing `tools` fails the prefix check.
6. Re-run your evals with an effort sweep that starts at the default, `high`. The docs reserve `xhigh` for the hardest long-running coding and agentic work, and `max` for measuring the ceiling.

In Claude Code, the migration guide suggests running `/claude-api migrate this project to claude-fable-5-2` to let the agent make the code changes [[5]](https://platform.claude.com/docs/en/models/fable-5-2/migration-guide).

## What are the limits and open questions?

The main limit is cost. At $8 and $40, Fable 5.2 costs twice as much per token as Opus 5.5, and its default effort is higher [[2]](https://platform.claude.com/docs/en/models/fable-5-2/overview). Several of the tools that make an append-only history practical are still in beta, including compaction on demand and mid-conversation tool changes. So is server-side fallback [[3]](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2). The benchmark numbers are Anthropic's own, several of them on internal or partial test sets. The announcement does not say how the 1,000-agent Fermat run was set up in enough detail to reproduce it [[1]](https://www.anthropic.com/claude-fable-and-mythos-5-2). Mythos 5.2 stays limited to verified organizations, so most teams can only use Fable 5.2 and its safeguards.

## Frequently asked questions

### How much does Claude Fable 5.2 cost?

Claude Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 on Claude Fable 5.1. Batch requests cost half, and fast mode costs $16 and $80 per million tokens.

### What is the difference between Claude Fable 5.2 and Claude Mythos 5.2?

They are the same underlying model with the same specifications and prices. Fable 5.2 has cybersecurity and biology safeguards and is available to everyone, while Mythos 5.2 drops some of those safeguards and is limited to organizations verified through Anthropic's verification programs.

### Why does Claude Fable 5.2 return a 400 error about a different conversation?

The API's prefix check found that the system prompt, the tools or an earlier message changed after the model produced a thinking block you sent back. Keep the history append-only, or as a stopgap use the thinking-binding-controls-2026-08-01 beta header with prefix_mismatch_behavior set to drop_block.

### Does Claude Fable 5.2 support forced tool use?

No. Setting tool_choice to any or to a named tool returns a 400 invalid_request_error. Keep tool_choice on auto and use strict tool use or JSON outputs when you need schema-valid input.

### Do I need to change code if I use Claude Managed Agents or the Agent SDK?

Anthropic says nothing changes when Claude Code, claude.ai, Claude Managed Agents or the Claude Agent SDK builds your requests. For Managed Agents, update the model name, update the advisor model if the agent has one, and remove speed fast from the model object if it is set.

## Sources

1. [Introducing Claude Fable 5.2 and Claude Mythos 5.2](https://www.anthropic.com/claude-fable-and-mythos-5-2) — Anthropic, accessed 2026-11-10
2. [Claude Fable 5.2](https://platform.claude.com/docs/en/models/fable-5-2/overview) — Anthropic (Claude Platform Docs), accessed 2026-11-10
3. [What's new in Claude Fable 5.2](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2) — Anthropic (Claude Platform Docs), accessed 2026-11-10
4. [Refusals and fallback](https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback) — Anthropic (Claude Platform Docs), accessed 2026-11-10
5. [Claude Fable 5.2 and Claude Mythos 5.2 migration guide](https://platform.claude.com/docs/en/models/fable-5-2/migration-guide) — Anthropic (Claude Platform Docs), accessed 2026-11-10