The short version

Fable 5.2 has been out for two days: top of the benchmark table, 20% cheaper, and Copilot adopted it on day one. But Anthropic's official migration guide hides 4 breaking changes and 2 silent default changes. If you just swap the model ID and ship, your production traffic can start returning 400s.

The key point: none of these four changes make the model worse. They make the request fail outright. Run the checklist below before you upgrade.

1. Prefix checks are now enforced for everyone: editing history returns 400

Each thinking block is only valid for the system prompt, tools, and history it was issued with. Fable 5.1 enforced this only for accounts created after 2026-08-31. Fable 5.2 enforces it for everyone.

The pattern old accounts used to get away with — rewriting an earlier turn mid-conversation and continuing — now returns 400 Invalid signature in thinking block, and retrying does not help.

The correct posture is append-only history. Test whether your pipeline secretly rewrites history with the test header first:

curl https://api.anthropic.com/v1/messages \ -H "anthropic-beta: thinking-binding-controls-2026-08-01" \ -d '{"model": "claude-fable-5-2", "prefix_mismatch_behavior": "drop_block"}'

Then turn on input_transformations logging and find which step is rewriting history.

2. Older models can't read Fable's thinking blocks

Routing, retries, and refusal fallbacks are standard agent plumbing — but Fable 5.1 and Opus 5.5 cannot read Fable 5.2's thinking blocks. When a fallback happens, the API silently drops those blocks. The request succeeds, but the next turn has lost its reasoning context, and cost and latency both go up.

Do this instead:

  1. Only write reasoning effort into Fable 5.2 messages — never into Fable 5.1 messages.
  2. Don't migrate history. A user session moving from Fable 5.1 to Fable 5.2 starts fresh.
  3. Tell users plainly: switching models loses the conversation context.

3. computer_20251124 is gone

The old computer tool version field computer_20251124 is retired. The replacement is computer_toolset_20260801. Skip the upgrade and you get a 400.

Two details that trip people up:

  1. The API tool name is computer_use_20260201; the beta tool name is computer-use-2026-01-14.
  2. computer_use_20260201 and computer_use_20260211 are two separate versions — fill them in separately.

4. The advisor must be from the same family

This is the easiest of the four to miss: the advisor must be from the same model family as the main model.

Fable 5.2 only pairs with Fable 5.2 or Mythos 5.2 (beta name mythos-advisor-5-2-20260115). Using Opus 4.8 as the advisor returns a 400 directly.

Three details:

  1. In the request, set both the advisor's model and advisor_model fields to the same model.
  2. Mythos 5.2 is not publicly released yet — for now, Fable 5.2 as advisor is the safest choice.
  3. After changing the advisor config, re-run the full test pass.

5. Two silent changes: the defaults moved

Fable 5.2 changed two defaults without telling you at upgrade time.

  1. tool_choice changed from auto to required. Under Fable 5.1 you could leave tool_choice unset and everything worked; under Fable 5.2, leaving it unset means the model decides for itself whether to call a tool — and you get "should have called a tool but answered directly" behavior. The fix: set tool_choice explicitly to auto.
  2. The default effort moved up a tier, from low to medium. The price math: Fable 5.2 is 20% cheaper than 5.1, but if your task only needed low effort, the higher default spends the discount right back. The fix: pin effort explicitly — low for cheap tasks, high for hard ones.

6. The cost math: don't default to the top tier

The official numbers: Fable 5.2 is $8/$40 per million tokens, down 20% from 5.1's $10/$50. Opus 5.5 is $4/$20 — half the per-token price. Per token, Fable is the expensive one; per task, it wins on long, hard work, because fewer steps mean fewer total tokens.

The trap is the effort tier. Artificial Analysis scores Fable 5.2 at 62 on High effort vs 65 on Max — 3 points apart — while the Max tier runs about $26.50 per task, roughly 2.6x Opus 5.5, with a 216-second first token. High effort gets you nearly the same score for far less money and time.

And the rate limit is separate: Fable 5.2's limit is independent of Fable 5.1's. Confirm your quota before you switch traffic, or you will get throttled on launch day.

Pre-launch checklist 

1. Pin the model ID to claude-fable-5-2 and pin effort explicitly 2. Confirm the workspace has 30-day data retention enabled 3. History is append-only; verify with the drop_block test header 4. Computer tool moved to computer_toolset_20260801 5. Advisor changed to Fable 5.2 / Mythos 5.2 6. tool_choice set to auto, with an explicit instruction 7. Handle stop_reason refusal; fallback targets are Opus 4.8 / Opus 5 only 8. Confirm Fable 5.2's independent rate limit 9. Re-run your own eval per effort tier before cutting over traffic

The honest limitations

  • Every benchmark figure above is vendor-reported by Anthropic (Artificial Analysis is the independent cross-check); nobody had independently reproduced the launch numbers.
  • Mythos 5.2 is not publicly released yet — anything about it, including the advisor beta name, comes from the migration guide, not from a public launch.
  • The four breaking changes and the two default changes come from Anthropic's official migration guide as summarized on 2026-11-10; verify against your own account behavior before trusting them in production.

Which model are you running, and which of these 9 checklist items are you stuck on? Tell me in the comments and I'll help you think through the migration path.

Adapted from the Chinese original published on Juejin on 2026-11-11 ("Fable 5.2 迁移避坑：4 个会让你的代码报 400 的改动"). Same facts, restructured for dev.to readers. Facts checked against Anthropic's migration guide and the November 9 announcement on 2026-11-10/11.