Claude Fable 5.2: Should You Switch From Fable 5.1 or Opus 5.5?

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026. According to Anthropic's Fable 5.2 migration guide, Fable 5.2 costs $8 / $40 per 1M input / output tokens, down from $10 / $50 for Fable 5.1. It is available on the Claude API, Amazon Bedrock, Claude Platform on AWS, Claude Models on Google Cloud and Microsoft Foundry.

Whether you should switch depends on where you start. From Fable 5.1, Anthropic calls the move mostly drop-in, and it is 20% cheaper. From Opus 5.5, Fable 5.2 costs twice as much, requires 30-day data retention and runs at a higher default effort, which raises cost without any error.

This article covers Anthropic's reported results, what breaks on each path, the silent effort change, and what happens to Fable 5.2's reasoning when a router or fallback moves a conversation to another model.

What Fable 5.2 brings

Fable 5.2 keeps a 1M-token context window and up to 128K output tokens per request. Adaptive thinking is always on, and the effort parameter has five levels: low, medium, high, xhigh and max.

Claude Mythos 5.2 has the same capabilities and pricing. It is available only to organizations verified through Anthropic's verification programs, such as the Cyber Verification Program, so it is not an option for most teams.

The gains Anthropic reports are mostly on coding and long agentic work. Here are the two headline results from its system card.

These are Anthropic's numbers, not independent replications. On the Remote Labor Index, 41.3% is the share of freelance projects Fable 5.2 completed at least as well as a human professional. The system card summary does not give the effort level or harness used for each model.

Fable 5.2 does not win everywhere. Anthropic says it scored highest on 14 of the 16 results in its capability table, trailing Opus 5.5 on the GDPval-AA and AA-Briefcase Elo ratings. None of these scores tell you the cost per completed task in your application.

For background on how the two families compared before, see our Claude Opus 5 vs Claude Fable 5 benchmark and our coverage of the return of Claude Fable 5.

Coming from Fable 5.1: switch after four checks

Anthropic describes this migration as mostly drop-in. The API surface, always-on adaptive thinking, the forced tool choice restriction, refusal handling, stop_details categories, the 30-day retention requirement and the high default effort all match Fable 5.1.

The price case is simple. Fable 5.2 costs $8 / $40 per 1M input / output tokens, against $10 / $50 for Fable 5.1. Cache reads drop from $0.25 to $0.20 per 1M tokens, according to Anthropic's pricing page.

The prefix check matters if your own code builds the messages array. A Fable 5.2 thinking block is valid only against the system prompt, tools and history that preceded it. Fable 5.1 enforced this only for accounts created on or after August 31, 2026, so older accounts can see new 400 errors.

Nothing changes if Claude Code, claude.ai, Claude Managed Agents or the Claude Agent SDK manages your history. Otherwise, Anthropic suggests running a session with the thinking-binding-controls-2026-08-01 beta header and prefix_mismatch_behavior: "drop_block", then fixing every prefix_binding_mismatch it logs.

Fable 5.2 also has its own rate limit, separate from the one Fable 5.1 and Fable 5 share. Check it before you move production traffic.

Coming from Opus 5.5: a workload decision

Fable 5.2 and Opus 5.5 use the same Messages API and tool use patterns. Both always think adaptively and reject forced tool choice, so those parts of your requests carry over.

Anthropic's own advice leans toward Opus. Its What's new page says to start most workloads on the latest Opus model, and to move to Fable 5.2 when your evals on Opus at higher effort still fall short.

Retention is the first gate. Fable 5.2 requires 30-day retention and isn't available under a zero data retention (ZDR) arrangement unless Anthropic expressly authorizes it. On the Claude API, every request from an organization or workspace without 30-day retention returns a 400 error. The rule is documented in Anthropic's model-specific data retention requirements.

Fast mode keeps the same ratio. It gives up to 2.5× higher output speed, at $16 / $80 per 1M input / output tokens for Fable 5.2 against $8 / $40 for Opus 5.5.

The silent cost change only hits Opus migrations

Opus 5.5 defaults to medium effort, while Fable 5.2 defaults to high on the API. A request that omits effort runs one level higher after the switch. It still succeeds, so the change shows up only in cost, latency and results.

Teams on Fable 5.1 are not affected: that model also defaults to high.

Three Fable 5.2 behaviors can raise token use further, on both migration paths:

• Prompt shortcuts stop working: in Anthropic's testing, lines such as "answer directly without deliberating" didn't stop Fable 5.2 from thinking at medium effort and above. Remove them and lower the effort level instead.
• High-effort sessions can add reviewers: at xhigh and max, Anthropic observed Fable 5.2 starting subagents to review its own work. Anthropic says subagents can account for a large part of a session's cost.
• Thinking shares the output limit: thinking tokens count toward max_tokens, so a small value can end the request after thinking with no text. Raise max_tokens or lower effort.

Who should switch?

Routing and fallback: where Fable 5.2's reasoning gets lost

On the Claude API and Claude Models on Google Cloud (Google managed), Fable 5.2 reads thinking blocks from Opus 5.5 and Fable 5.1. A conversation that moves onto Fable 5.2 keeps its earlier reasoning. The reverse does not work: no model other than Fable 5.2 and Mythos 5.2 reads Fable 5.2's blocks.

When a router switch, a client-side retry or a fallback moves a conversation off Fable 5.2, the API drops the unreadable blocks. The request succeeds and the dropped tokens aren't billed. Anthropic warns that the next model then works without that reasoning, which can raise cost and latency on the first turn after the switch.

The same applies to Anthropic's own classifier refusal fallback. With fallbacks: "default" (beta), declined requests retry on Opus 4.8 or Opus 5, without Fable 5.2's thinking. Requests declined for reasoning_extraction are not retried. Our Claude Opus 5 benchmarks show what that fallback target can do. Fable 5.2 runs the same stop_details categories as Fable 5.1, so Opus 5.5 users should handle every category, not only the ones they have seen.

Comparing the models with Eden AI

Eden AI gives you one API key and one OpenAI-compatible endpoint for more than 1,000 models from more than 80 providers. Comparing two models on the same requests then comes down to changing the model ID.

You can choose the order of providers, or sort them by price, speed or latency, and set up automatic fallback to backup models. A comparison run looks like this:

1. Set effort explicitly on every request. Test medium and high before higher levels.
2. Replay a representative mix of short requests and long agent sessions on both models.
3. Record completed tasks, total tokens, latency and retries.
4. Test a live conversation that falls back from Fable 5.2 to another model.
5. Move only the workloads where the measured gain covers the added cost and constraints.

Treat fallback as a way to keep the service up, not to keep the reasoning. If a fallback chain moves a conversation from Fable 5.2 to another model, the backup model works without Fable 5.2's earlier thinking, so expect the first turns after the switch to cost more and run slower.

What to keep in mind

• The benchmarks are Anthropic's: reproduce the comparison with your own prompts, tools and acceptance criteria before moving traffic.
• Retention rules depend on the platform: Anthropic's retention page covers services where Anthropic processes the data. On Amazon Bedrock and Google-managed Google Cloud, the cloud provider's documentation applies.
• Higher effort is not automatically better: Anthropic says most of the gain is already present at high, and max runs slower and costs more, often for a small gain. Anthropic publishes no figure for subagent cost.
• Agent safety under fallback: Anthropic's system card reports that Fable 5.2 falls back to Opus 4.8 more often than recent models in prompt injection scenarios, resulting in a higher attack success rate.
• No Priority Tier: neither Fable 5.2 nor Mythos 5.2 is supported on Priority Tier.

FAQ

Further reading

• Anthropic's Fable 5.2 migration guide: the breaking changes and checklists for each path, from Fable 5.1 and from Opus 5.5.
• Claude Fable 5.2 system card: Anthropic's capability and safety evaluations, including the benchmark results above.
• Anthropic's effort documentation: default levels per model and how to run an effort sweep.
• Claude Opus 5 vs Claude Fable 5 benchmark: how the two model families compared one generation earlier.
• Claude Fable 5 vs GPT-5.5 benchmark: Fable against a non-Anthropic alternative, for broader model selection.