On November 9, 2026, Anthropic released Claude Fable 5.2 (claude-fable-5-2), available on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. Claude Mythos 5.2 is the same model with fewer blocks around offensive security and biology, open only to organizations approved through Anthropic's verification programs. Both cost US$8 per million input tokens and US$40 per million output tokens, down from US$10 and US$50 for version 5.1. The migration guide calls the switch "mostly drop-in", then lists four changes that can break production code.
What changes in the API
The change with the widest reach is the prefix check. When the model reasons, the API returns thinking blocks that the application sends back on later calls in the conversation. On Fable 5.2, those blocks are only valid if everything before them is unchanged: the system prompt, the set of tools and every earlier message. If the application changed any of those between one call and the next, the response is a 400 invalid_request_error.
The documentation lists the most common edits that trigger the error: rebuilding the system prompt with today's date or a mode flag, re-rendering context in the first message, shortening old tool results to save tokens, reordering messages, and adding or removing tools mid-conversation. These are common techniques for controlling cost and context, and many integrations use at least one of them.
The rule already existed, with exceptions. On Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5, accounts created before August 31, 2026 did not get the error: invalid blocks reached the model and the call went through. On Fable 5.2 and Mythos 5.2 the check applies to every account. The documentation itself warns that a run with no errors on your own API key does not prove your code keeps the prefix intact, so testing with an older account is not enough.
The other three changes affect fewer teams:
- anyone using the computer use tool in its
computer_20251124version gets a 400 error on every platform except Bedrock and needs to move tocomputer_toolset_20260801; - with the advisor tool (in beta), a Fable 5.2 executor only accepts Fable 5.2 or Mythos 5.2 as its advisor;
- Fable 5.1 cannot read 5.2's reasoning blocks, so a conversation that moves back to the older model loses that reasoning, without an error.
Who is affected
The most exposed group is companies with their own product calling the Messages API with reasoning enabled and running long conversations: in-product assistants, support agents, automations that chain tools. Teams on Claude Managed Agents only need to change the model name, according to the guide, plus the advisor's model when there is one.
Data retention is another factor. Fable 5.2 requires 30-day retention, like 5.1, and does not run under a zero data retention arrangement without express authorization from Anthropic. For teams coming from Opus 5.5 under zero retention, this is the first barrier: the API returns a 400 error to every call from an organization or workspace without the 30 days. That is a contract and legal decision before it is an engineering task.
Price also depends on where a team starts. For current Fable 5.1 users, 5.2 is 20% cheaper per token. For Opus 5.5 users, who pay US$4 and US$20, Fable 5.2 costs twice as much. Anthropic still recommends Opus 5.5 as the everyday model. The gains it publishes for 5.2 on agentic coding tests (69.9% on SWE v2 against 62.3% for Opus 5.5) are the vendor's own numbers, measured under its own conditions.
What to do before switching models
There is no deadline. Anthropic has not announced a retirement date for Fable 5.1 or Opus 5.5, so teams that do not plan to switch have nothing to do now. For teams that will switch, the migration guide's checklist comes down to four steps, with retention settled first:
- If the organization has zero retention, sort out 30-day retention with Anthropic or set up a separate workspace before any testing.
- Find the places in the code where the system prompt, the tool list or old messages change during a conversation, and make the integration append-only: history grows at the end and nothing earlier is rewritten.
- Update computer use and the advisor's model, if the integration uses those tools.
- Run the test suite against Fable 5.2, because the account's old behavior is not a reference.
There is a temporary escape hatch: the prefix_mismatch_behavior: "drop_block" option, in beta, discards invalid blocks and lets the call through. The documentation treats it as a stopgap, because the model loses its earlier reasoning and tends to spend more output tokens redoing it. On the bill, that can cancel out the price cut.
Switching models has an engineering cost even when the vendor stays the same, and that cost belongs in the comparison, as I discussed in The cost of a tool includes the work to replace it. Here, the work is concentrated in how the application builds the conversation history.
What is still unknown
Anthropic has not given a retirement date for Fable 5.1, nor said whether the all-accounts prefix check will extend to earlier models. There are no independent numbers on 5.2's performance, or on how much drop_block mode raises consumption in real conversations. Access to Mythos 5.2 depends on case-by-case approval, with no public criteria beyond the verification programs.
If your company is weighing something similar, reach me on LinkedIn or WhatsApp.
