Anthropic shipped Claude Fable 5.2 on Monday, November 9. It's cheaper than 5.1 — $8 per million input tokens and $40 per million output instead of $10 and $50 — and it took the top spot on the Artificial Analysis Intelligence Index the same day. Model ID: claude-fable-5-2.

So the obvious move is swapping one string and redeploying. I did the boring version first: I read Anthropic's official migration guide, then grepped every repo on my machine for the patterns it warns about. The guide is blunt. Moving from Fable 5.1 is "mostly drop-in" — and then it lists four changes that can break working code. Not behavior shifts. Breaks.

1. The prefix check now applies to every account

Every thinking block Fable produces is only valid against the system prompt, tool list, and conversation history that came before it. If your code edits an earlier turn — rewrites the system prompt, swaps a tool, trims history — and then sends a thinking block back, the API returns a 400.

On Fable 5.1, that check only ran for accounts created on or after August 31, 2026. On 5.2, it runs for everyone, whatever the account's age. The guide says it plainly: "on an older account, requests that pass on Claude Fable 5.1 can fail on Claude Fable 5.2."

One nasty detail: retrying the same request body doesn't clear the error. And if you use the MCP connector without pinning a server's tool list, a tool list that changes between requests counts as an edit for 5.2 thinking blocks — even though 5.1 and Opus 5.5 blocks don't care.

The fix the guide suggests is a check session with the thinking-binding-controls-2026-08-01 beta header and prefix_mismatch_behavior: "drop_block": log the input_transformations entries and fix every prefix_binding_mismatch. In production, leave the behavior at the default "error" if a mismatch can only mean a bug. drop_block is a stopgap — dropping blocks can raise output tokens, since the model re-derives the reasoning it lost.

2. Older models can't read Fable 5.2's thinking blocks

Fable 5.2 can read Fable 5.1's thinking blocks, so moving a conversation upward keeps its reasoning. Going the other way, Fable 5.1 can't read 5.2's blocks. If a router, a client-side retry, or a classifier refusal fallback moves a conversation from 5.2 down to 5.1 or Opus 5.5, the API drops the blocks the new model can't read. The request still succeeds, and you aren't billed for the dropped input tokens. But the model then works without that reasoning, and the guide says to expect higher cost and latency on the first turn after the switch.

The permitted fallback targets for 5.2 are Claude Opus 4.8 and Claude Opus 5. If your fallback chain points at 5.1 or Opus 5.5, audit it before you switch the model string.

3. The old computer use tool is gone (except on Bedrock)

A tools entry of type computer_20251124 returns a 400 on every platform except Amazon Bedrock. The replacement is the computer_toolset_20260801 toolset — declared with no name and no display dimensions, and with the beta header dropped. Straight grep: if that string appears anywhere in your codebase, you're broken on 5.2 outside Bedrock.

4. Advisors have to match the executor

With the beta advisor tool, a Fable 5.2 executor accepts only a Fable 5.2 or Mythos 5.2 advisor. Anything else — including Fable 5.1 — returns a 400 invalid_request_error.

The trap inside the trap: changing the advisor's model is itself an edit to tools, so if you make the change mid-conversation and then send earlier thinking blocks back, you trip the prefix check from item 1 and get a 400 for a different reason. The guide says to make the change in new conversations.

What my own code check actually found

Here's the honest part. I grepped my whole workspace — every .ts, .js, and .py file, excluding node_modules — for computer_20251124, tool_choice, budget_tokens, max_tokens, and any direct Anthropic SDK usage. Nothing. Not one file on this machine calls the Claude Messages API directly. The only matches for "claude-fable" anywhere were my own agent's session logs talking about the migration guide.

So my migration is, in the guide's words, a model-picker change: if Claude Code, claude.ai, Managed Agents, or the Agent SDK builds your requests, changing the model ID is the whole migration. That's most developers — and the four breaking changes above only bite if your code builds the messages array itself.

But "my code has nothing to break" is a finding, not an excuse to skip the check. The grep took about a minute. Search for: computer_20251124, any tool_choice of any or a named tool (both return 400 on 5.2), thinking: {type: "disabled"} or budget_tokens (both 400), and any place you edit messages between turns instead of appending.

Three quieter changes worth knowing

First, prompt lines that ask for less thinking — "answer directly without deliberating" — don't stop Fable 5.2 from thinking at medium effort and above. On 5.1 they mostly worked. On 5.2, remove those lines and lower the effort level instead. Related: a small max_tokens can now run out during thinking before any text appears, because thinking tokens count toward the limit.

Second, at xhigh and max effort, Anthropic observed 5.2 spinning up subagents to review its own work. 5.1 didn't do that on the same tasks. Subagents can be a large part of what a session costs. Third, the effort default is high on the API (medium in Claude Code).

The 30-day retention 400

Fable 5.2 requires 30-day data retention. On the Claude API, a request from an organization or workspace without it returns a 400 invalid_request_error on every request. Same as 5.1 — but if you never migrated to 5.1, this is your first encounter with it, and the error gives no hint that retention is the cause.

My checklist, stolen from the guide

  • Swap claude-fable-5-1 for claude-fable-5-2 (Bedrock: anthropic.claude-fable-5-2).
  • Grep for computer_20251124, forced tool_choice, disabled thinking, and budget_tokens.
  • Run the prefix check with drop_block in a test session and fix every mismatch before production traffic moves.

The price cut is real — 20% is 20%. But the guide's actual message is narrower than the launch headlines: the model got cheaper and the contract got stricter.

Sources: Anthropic, "Claude Fable 5.2 and Claude Mythos 5.2 migration guide" (platform.claude.com/docs/en/models/fable-5-2/migration-guide), read November 10, 2026. The four breaking changes and the retention rule are from that guide; the grep results are from my own machine, run the same morning.