Claude Fable 5.2 Migration: API 2026-10-21 and the 400 Errors

In what order should you ship this migration?

Ship one change at a time, because a combined release makes any 400 hard to attribute. Anthropic's documentation sets the condition for the last step: move to Fable 5.2 when your evals on a Claude Opus model at higher effort still fall short. Claude Code users can run /claude-api api-version, /claude-api preserved-thinking-migration and /claude-api migrate this project to claude-fable-5-2 for much of the code work.

Which error are you looking at?

Match your message here, then read the matching section. The first row has two separate causes: the API version rejects these parameters outright, and many recent models reject non-default values on any version.

What does API version 2026-10-21 actually remove?

The hard case is low temperature. Anthropic's guide says no other setting replaces it. If a classifier on a model such as Claude Sonnet 4.6 depends on a low value, you can stop sending it or keep that request on 2023-06-01. Treat the second option as a holding position while you build an eval, since current models reject non-default values anyway.

• •Documents: a document block directly in an assistant message returns a 400. Send it in a user message.
• •Citations: a text block with a non-empty citations array in a user message returns a 400.
• •Admin and Analytics APIs: roles becomes role_ids, group_id becomes rbac_group_id, actor.email becomes actor.email_address, summaries becomes data. On API keys, workspace_id is replaced by scope, with different values.
• •Types: MessageParam is now a union per role. A TypeScript interface can no longer extend it, so use a type alias.
• •Removed: compactionControl in the TypeScript and Ruby tool runners (use server-side compaction, with a trigger of at least 50,000 tokens), and Text Completions in TypeScript, Go, Java, Ruby and the CLI.
• •Packaging: move TypeScript platform packages to v1 in the same upgrade, or you get a second, older core SDK. The old Go path for the ant CLI keeps installing v1 with no warning.

Why can Authorization: Bearer break a gateway that worked yesterday?

The Claude API accepts a key in either header, so direct calls keep working and the failure shows up in whatever sits between you and Anthropic. Anthropic names middleware, log filters and test stubs as well as gateways. Check any log filter that redacts X-Api-Key, because the key now travels in a different header. If you own the proxy, change it to read Authorization. If another team owns it, send X-Api-Key yourself through the SDK's documented hook until they move.

• •Unchanged: Foundry clients still send x-api-key, and Bedrock Mantle clients already sent a bearer token.
• •New rule: a client holds one credential. Passing two raises ValueError in Python and TypeError in TypeScript.

What is the Bedrock trap in the new SDKs?

On Bedrock, 2026-10-21 exists only for Claude 4.7 and later, and Anthropic tells anyone calling an older model to stay on their current SDK. A user report filed on 22 October, open with no maintainer reply when we read it, shows the failure: AnthropicBedrock 2.x is rejected for Sonnet 4.5, Opus 4.5 and Opus 4.6, while the same code works on 1.17.0. Follow the documented route and pin the SDK. Use the reporter's override only if one codebase must mix old and new Bedrock models.

What are the four Claude Fable 5.2 breaking changes?

Anthropic calls the move from Fable 5.1 mostly drop-in, with four exceptions. Coming from Opus 5.5, the retention rule and default effort are also new. Coming from Opus 5 or earlier, forced tool choice and always-on thinking break too.

• •The prefix check applies to every account. On Fable 5.1 it errored only for accounts created on or after 31 August 2026, so history-editing code can pass today and fail after the switch.
• •Only Fable 5.2 and Mythos 5.2 read Fable 5.2 thinking blocks. Moving to any other model loses that reasoning: the request succeeds, dropped blocks are unbilled, and the first turn after the switch can cost more and run slower.
• •computer_20251124 returns a 400 everywhere except Bedrock. Send one computer_toolset_20260801 entry with no name or display dimensions, drop the beta header and update your agent loop.
• •A Fable 5.2 executor needs a Fable 5.2 or Mythos 5.2 advisor. Changing the advisor is itself an edit to tools, so apply it in new conversations.

Why does Claude say the block is bound to a different conversation?

Each thinking block is valid only against the system prompt, tools and messages before it. Change any of those and that block, plus every later one, fails. Effort, max_tokens, tool_choice and cache_control are outside the check. A clean test run can mislead: on an older account Opus 5.5 lets failing blocks through silently, so your key passes while a customer's newer key gets the 400. In Message Batches an item with the field unset has the failing blocks dropped with no error. Anthropic's rule is to keep system and tools fixed and only add to the end of messages. Nothing changes if Claude Code, claude.ai, Claude Managed Agents or the Claude Agent SDK manages your history. One user report from September describes Opus 5.5 chats through Bedrock in Cursor failing permanently with this error after long turns; Cursor staff confirmed the failures.

• •400 by default on every account: Fable 5.2, Mythos 5.2 and, on the Claude API, Mythos 5.1.
• •400 by default only for accounts created on or after 31 August 2026: Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5.
• •Changing system prompt (a date, a mode flag): use a mid-conversation system message.
• •Adding or removing tools: use tool_addition and tool_removal blocks, and pin MCP connector tool lists.
• •Shortening old tool results: shrink them before first send, or trim on the server.
• •Client-side summarising that keeps recent thinking: use on-demand compaction.
• •Saved sessions rebuilt from templates: store exactly what you sent and replay that.

Which changes succeed quietly and still cost you?

The hardest regressions return HTTP 200. Fable 5.2 lists at $8 input and $40 output per million tokens against $4 and $20 for Opus 5.5. As illustrative arithmetic only, 1,000 requests of 20,000 input and 2,000 output tokens is 20 million in and 2 million out, before caching. Opus 5.5 costs (20 x $4) + (2 x $20) = $120, about Rs 11,263 at the RBI reference rate of Rs 93.8579 per USD on 9 November 2026. Fable 5.2 costs (20 x $8) + (2 x $40) = $240, about Rs 22,526. Cost per finished task is covered in /blog/claude-fable-5-2-vs-opus-5-5-cost-per-task-rupees.

What has to be finished by 30 November 2026?

Both removals apply whether or not you adopt Fable 5.2. Budget for the first properly: Sonnet 5.5, Anthropic's recommended replacement, rejects forced tool choice, sampling parameters and thinking budgets, so a routine-looking swap touches the same code as this guide.

• •Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) retires on the Claude API.
• •Fast mode on claude-opus-4-8 is removed. Those requests will return an error and, per the Claude Code changelog, will not fall back to standard speed. Anthropic points to Opus 5.5 for fast mode.

Should you adopt message threads during this migration?

Message threads, in beta since 2 November, keep the conversation on the server so later requests carry only new messages. They help time to first token in long loops and do not change price or rate limits. Finish the append-only work first, then pilot them on your longest loops. A thread is best-effort: it expires 5 minutes after its last request, and a continue can return a 404 with thread_not_found, so you must keep your own copy and resend the full conversation. Threads do not work with Message Batches, token counting or compaction, nor for HIPAA-ready organisations or those with inference hooks, and they are not eligible for zero data retention. Fable 5.2's retention rules are covered in /blog/fable-5-2-30-day-retention-inference-hooks-dpdp-claude-code-security.

What is not yet known?

Fable 5.2 is two days old at the time of writing. These gaps affect planning.

• •How much drop_block raises output tokens on your workload. Anthropic says it can be significant and tells you to measure it.
• •A retirement date for 2023-06-01. We found none in the migration guide or release notes.
• •Whether the Bedrock report and the Cursor thread have been resolved. Neither showed a released fix when read, though a commit referencing the Bedrock report exists.

The migration checklist

Copy this into the ticket and strike what does not apply.

Where Tech Arion can help

Tech Arion is a Hyderabad software and AI consultancy that builds with Claude Code. If your product assembles its own Claude requests, we can audit history handling, fix gateway and SDK changes, and set up the eval and effort sweep before traffic moves.

Frequently asked questions about migrating to Claude Fable 5.2

Short answers to the questions engineers ask first.