Migrating from Claude Opus 5.5 to Fable 5.2: when and how teams should move

Opus 5.5 vs Fable 5.2, side by side

Every number below comes from Anthropic's models overview, Fable 5.2 model page and migration guide.

Benchmarks Anthropic published

From the launch announcement. These are Anthropic's own runs. The announcement also notes the models voice suspicion that they are being evaluated more often than recent models, which can make evals less reliable. Your evals on your tasks matter more than this table.

Where Fable 5.2 runs: Bedrock, Google Cloud (Vertex AI), Foundry

Check this before anything else if you buy Claude through a cloud provider. Anthropic's model page lists Fable 5.2 on the Claude API, Claude Platform on AWS, Microsoft Foundry, and Claude Models on Google Cloud (both the Google-managed and the Anthropic-managed options). The ID is claude-fable-5-2 on all of them.

• Amazon Bedrock. The models overview shows a dash in the Bedrock ID column for Fable 5.2 at launch, while Opus 5.5 has anthropic.claude-opus-5-5. Anthropic has not published a Bedrock ID for Fable 5.2 yet. If your stack is on AWS, Claude Platform on AWS is the documented route today. Check the Bedrock console before you plan around it.
• Google Cloud (Vertex AI). Listed, with the same ID. Fast mode works there with speed: "fast", and on the Google-managed option Google sets the prices.
• Microsoft Foundry. Listed. Fast mode is a separate deployment there, claude-fable-5-2-fast, and you leave out speed.

When to switch and when to stay

Anthropic's own docs tell most teams to start on Opus 5.5. Fable 5.2 is for "demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still fall short." I'd read that as: move workloads, not the whole company.

Step-by-step migration

Eight steps in the order we'd run them. Anthropic also ships a shortcut: in Claude Code, /claude-api migrate this project to claude-fable-5-2 rewrites the code for you. You still own the evals, the cost check and the rollout.

1. 1. Clear data retention first

   Fable 5.2 requires 30-day data retention. If your organization or workspace runs Opus 5.5 under a zero data retention arrangement, every Fable 5.2 request on the Claude API returns a 400 invalid_request_error. Enable 30-day retention for a workspace, or talk to your Anthropic account team, before anyone touches code. Anthropic lists this as the first item on its Opus 5.5 checklist.

2. 2. Swap the model ID behind a flag

   The ID is claude-fable-5-2 on the Claude API, Claude Platform on AWS, Microsoft Foundry and both Google Cloud options. The Messages API and tool use patterns are the same as Opus 5.5, so the change is one string. Put it behind a config value or feature flag so you can route a share of traffic and flip back without a deploy.

   # Before
   model = "claude-opus-5-5"
   
   # After
   model = "claude-fable-5-2"

3. 3. Set effort explicitly

   Opus 5.5 defaults to medium effort. Fable 5.2 defaults to high. A request that omits effort still succeeds, but it now thinks more, so cost and latency move with no error to warn you. Set output_config.effort on every request, start where Opus 5.5 ran, then sweep levels on your evals. Thinking tokens count toward max_tokens, so a tight limit can stop after the thinking block with no text.

   curl https://api.anthropic.com/v1/messages \
     -H "x-api-key: $ANTHROPIC_API_KEY" \
     -H "anthropic-version: 2026-10-21" \
     -H "content-type: application/json" \
     -d '{
       "model": "claude-fable-5-2",
       "max_tokens": 16000,
       "messages": [{"role": "user", "content": "Review this migration plan"}],
       "output_config": { "effort": "medium" }
     }'

4. 4. Check prompts and conversation handling

   Two things break quietly. First, the prefix check: Opus 5.5 enforces it only for accounts created on or after August 31, 2026, while Fable 5.2 enforces it for every account. If your code builds the messages array itself and edits earlier turns, requests that pass today can return a 400 on Fable 5.2. Anthropic's guide shows how to run the check against Opus 5.5 first, using the thinking-binding-controls-2026-08-01 beta header. Second, Anthropic reports that prompt lines asking Fable 5.2 to skip deliberation don't stop it thinking at medium effort and above. Delete those lines and lower effort instead. Review the rest of your system prompts against Prompting Claude Fable 5.2.

5. 5. Re-test tool use

   Forced tool choice (any or a named tool) returns a 400 on both models, so that carries over. What changes: if you use the advisor tool (beta), a Fable 5.2 executor accepts only claude-fable-5-2 or claude-mythos-5-2 as its advisor. In an organization with customer-managed encryption keys, Fable 5.2 has no structured outputs, so remove strict: true and output_config.format and validate output in your code. Fable 5.2 also runs safety classifiers. Handle every stop_details.category, and decide whether declined requests retry server-side (fallbacks: "default", beta) or in your own code.

6. 6. Run your evals side by side

   Anthropic's own advice in the models overview is to start with Opus 5.5 and use Fable 5.2 when your evals on Opus 5.5 at higher effort still fall short. So the eval run is the decision. Replay the same task set on claude-opus-5-5 at its best effort and on claude-fable-5-2 at two or three effort levels. Score pass rate, tokens per task and wall-clock time. If Opus 5.5 at high or xhigh already passes, you may not need Fable 5.2 for that workload.

7. 7. Do the cost check with real token counts

   List price doubles: $8 and $40 per million input and output tokens against $4 and $20. Cache reads stay at $0.20 per million on both. Higher default effort can add output tokens on top. Take tokens per task from your eval run, multiply by the new prices, and compare cost per successful task, not cost per call. A model that finishes in fewer retries can cost less per outcome. Batch API requests are 50% off on both models. Also check rate limits: Fable 5.2 has its own limit, separate from Opus 5.5, with lower token limits at each tier.

8. 8. Roll out in stages

   Start with internal users or one low-risk workflow, then move to 5%, 25% and 100% of the target traffic. Hold each stage long enough to see error rates, refusal categories, latency and spend per task. Route new conversations, not live ones: Opus 5.5 can't read Fable 5.2's thinking blocks, so moving a conversation back mid-thread drops that reasoning.

Claude Code settings changes

Claude Code ships a fable alias. Outside Claude apps gateway sessions it resolves to Fable 5.2, unless you set ANTHROPIC_DEFAULT_FABLE_MODEL. Fable is not the default on any plan, so nobody moves until they pick it. Depending on plan and seat, Fable usage can bill to usage credits instead of plan limits, and the /model picker shows "Requires usage credits" when it does.

# Fable 5.2 needs Claude Code v2.1.321 or later
claude update

# Start a session on Fable 5.2
claude --model fable

# Or switch inside a running session
/model fable

# Behind a gateway that serves it, pick the full ID
/model claude-fable-5-2

To make it the team default, set the model field in a shared settings file. Admins can limit choices with availableModels.

{
  "model": "fable"
}

One detail worth knowing: the Claude Code docs say Fable 5.2 sessions default to medium effort there, even though the API default is high. Requests that Fable's safety classifiers flag fall back automatically: biology to Opus 5, cybersecurity to Opus 4.8.

Rollback plan

Rollback is the same one-line change in reverse, which is why step 2 puts the ID behind a flag. Write these down before stage one:

• Triggers. Pick numbers in advance: eval pass rate below the Opus 5.5 baseline, cost per successful task over budget, a rise in 400s or refusals.
• Keep the Opus 5.5 config. Keep its prompts and effort setting as a named config, not a deleted branch. Fable-specific prompt edits may not suit Opus 5.5.
• Expect a costlier first turn. Opus 5.5 can't read Fable 5.2 thinking blocks. The API drops them, doesn't bill them, and the request succeeds, but the next turn runs without that reasoning.
• Mind the dates. Anthropic commits to keeping Opus 5.5 until at least September 22, 2027, so it stays available to roll back to.

# API: one line back
model = "claude-opus-5-5"

# Claude Code: switch the session back
/model claude-opus-5-5

FAQ

Keep reading

• Claude Fable 5.2 for engineering teams: what changes for your team
• Hire engineers who already work with Fable 5.2
• Compare every Claude model on prices and specs
• Claude Fable 5.1 vs Opus 5, the previous round of this decision
• Claude Fable 5.1 pricing and benchmarks

Sources

• Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2
• Claude docs: Models overview
• Claude docs: Claude Fable 5.2
• Claude docs: Fable 5.2 and Mythos 5.2 migration guide
• Claude docs: Effort
• Claude Code docs: Model configuration