Claude Fable 5.2 and Mythos 5.2 Released With API Prices 20% Lower - Four Changes Can Break a Move From 5.1
Claude Fable 5.2 and Mythos 5.2, released November 9, cost $8 input and $40 output, 20% below 5.1. Four changes can break code, including 400 errors on older accounts when edited history returns thinking blocks.
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 20263. Per million tokens, the API charges $8 for input and $40 for output, which is 20% below the $10 and $50 of Fable 5.1 and Mythos 5.11. Every Claude API customer can use Fable 5.2, while Mythos 5.2 is limited to organizations that have passed vetting.
Updating the model ID may not be the whole job. The documentation names four changes from Fable 5.1 that can break code: three surface as 400 errors, and the fourth raises no error and simply loses reasoning1. Before putting the price gap into a budget, it is worth working out which of these apply to your integration.
A 20% Lower Rate, but the Bill Depends on Effort and Thinking
| Per million tokens | Fable 5.2 / Mythos 5.2 | Fable 5.1 / Mythos 5.1 |
|---|---|---|
| Input | $8 | $10 |
| Output | $40 | $50 |
| Cache reads | $0.20 | $0.25 |
Cache reads stay at 0.025 times the input rate, so all three lines fall by 20%12. In September, Fable 5.1 left input and output untouched and cut only cache reads, so this round differs in reaching input and output. On 5.2, batch work is $4 in and $20 out, and fast mode, which runs up to 2.5 times quicker3, is $16 in and $80 out.
Rates do not translate directly into invoices, though. Fable 5.1 usually obeyed a prompt line telling it to skip deliberation, but when Anthropic tested that kind of instruction on 5.2, it did not halt thinking at medium effort or higher, and from high upward the model reasons ahead of nearly every response1. Because thinking tokens are counted in max_tokens, requests depending on this kind of instruction to stay under a tight limit can be cut off before producing text. At settings like xhigh and max, Anthropic also saw Fable 5.2 spawn subagents to check its own output inside an agentic coding product, which 5.1 did not do on the same tasks, and subagents can be a large share of session cost2.
Effort defaults to high on the API1 and to Medium in the Claude apps and Claude Code. According to Anthropic’s announcement, early FrontierCode 2.0 testing had Fable 5.2 at High roughly matching the Max-effort score of Opus 5.5 (45.3% against 45.5%) while each task cost about half as much3. That is one evaluation reported by the vendor, and whether the ratio holds on your own tasks can only be established by measuring. The migration guide likewise asks teams to repeat an effort sweep on their own evals and take fresh cost and latency baselines2.
On model choice, the documentation advises beginning most work on the latest Opus and moving to Fable 5.2 when Opus at higher effort still falls short in your evals1.
The Four Changes That Can Break a Move From 5.1
Editing History and Returning Thinking Blocks Now Fails on Accounts of Any Age
When a thinking block generated by Fable 5.2 or Mythos 5.2 comes back to the API, the service verifies that the system prompt, tools, and prior messages ahead of it are the same as when it was generated. If this prefix check fails and prefix_mismatch_behavior is unset, the result is a 400 error1. Fable 5.1 enforces the error only for accounts dating from August 31, 2026 or later. Code that edits history on an older account can therefore run cleanly on 5.1 and start failing at the switch. Resending an identical request does not help2.
Requests assembled by Claude Code, claude.ai, Claude Managed Agents, or the Claude Agent SDK are unaffected by this item1. Hand-built requests are also safe when system and tools stay fixed through the conversation, messages is only appended to, assistant turns return exactly as the API sent them, and nothing on the official checklist applies.
For hand-built requests, the migration guide describes the test. Add the thinking-binding-controls-2026-08-01 beta header, set prefix_mismatch_behavior to "drop_block", run one session, and fix each prefix_binding_mismatch recorded in input_transformations. The work is needed even on accounts that have never seen this error on 5.1, and the setting lets any account perform the test2. Teams on the MCP connector without a pinned server tool list should point the test at claude-fable-5-2 as well, since a tool list that varies per request counts as an edit for 5.2 blocks while not counting as one for 5.1 or Opus 5.5 blocks.
Running "drop_block" in production discards the affected blocks and keeps requests moving, but the documentation frames it as a temporary measure until the edit is fixed and warns that output tokens can rise1.
Reasoning From 5.2 Does Not Carry Over to Other Models
This is the one item of the four that produces no error. Only Fable 5.2 and Mythos 5.2 can read their thinking blocks, so when a conversation moves to another model the API removes those blocks and the request still goes through1. The models that cannot read them include Opus 5.5, Fable 5.1, Mythos 5.1, and Fable 5.2’s fallback targets, Opus 4.8 and Opus 5. Removed tokens are not charged, but the receiving model has to carry on with the reasoning gone, and the next turn can end up slower and more expensive2.
In the opposite direction, conversations going from Fable 5.1 or Opus 5.5 to 5.2 keep what was reasoned so far, provided they run through the Claude API or Claude Models on Google Cloud (Google managed)1. The exposure is in setups where a router, retries in client code, or a refusal fallback can send a 5.2 conversation elsewhere. With the beta header, each such case is logged in input_transformations as model_binding_mismatch, which makes the frequency measurable. Replaying stored conversations under a different account has a similar effect: 5.2 thinking blocks are valid only for the account that created them or one linked to it2.
computer_20251124 Returns a 400 Outside Amazon Bedrock
Fable 5.1 accepted the older computer_20251124 tool. On 5.2, declaring it yields a 400 invalid_request_error on all platforms but Bedrock1. The replacement is computer_toolset_20260801: remove the beta header, leave out name and the display dimensions, and adapt the agent loop to the toolset2.
The Advisor Must Also Be a 5.2 Model
In the beta advisor tool, an executor running Fable 5.2 or Mythos 5.2 requires one of those same two models as advisor. Naming Fable 5.1 or Mythos 5.1 gets a 4001.
This ties back to the first item. Rewriting the advisor’s model counts as an edit to tools, so doing it in a running conversation and then returning earlier thinking blocks fails the prefix check unless "drop_block" is set. That is why the documentation asks for the change in fresh conversations. Teams can also leave the executor alone and move only the advisor to claude-fable-5-2, which every executor model accepts. Advice from a Fable 5.2 advisor arrives encrypted, however, so code that reads its text needs updating.
What to Check Beyond the Code
Rate limits are separate. Fable 5.2 draws on a dedicated allowance rather than the pool shared by Fable 5.1 and Fable 5, so check the ceiling ahead of moving production traffic. Mythos 5.2 stays in a single pool with Mythos 5.1 and Mythos 52.
During extended multi-stage work, Fable 5.2 may stop partway and hand back what is left or propose carrying on, and the documentation says this is more common at low and medium1. Teams lowering effort to save money should track completion rates alongside spend.
The 30-day data retention requirement matches 5.1, and zero data retention remains off the table unless Anthropic expressly authorizes it2. Access to Mythos 5.2 runs through vetting routes including the Cyber Verification Program, and the migration guide asks organizations to confirm they have access before changing model IDs.
Coverage of All Accounts Was Signaled in September
The restriction on editing history arrived with Fable 5.1 for new API accounts, along with a statement that future models would extend it to existing accounts. Anthropic describes it as a defense against distillation, and this release’s announcement confirms that it now covers all accounts3. In an October research post, the company itself listed the integration fixes some developers had to make as a side effect. The 5.1 announcement described new accounts created from that day on, while the current documentation sets the start at August 311. Integrations running Fable 5.1 on accounts created before that date may meet this error for the first time when they move to 5.2.
Retirement, at least, gives no reason to hurry for now. The model deprecations page shows Fable 5.1 as Active, with a tentative retirement “not sooner than” September 1, 2027 on Anthropic-operated platforms.
A sensible order starts with testing any code that builds its own history. Next come the computer use and advisor declarations, then a search for every path through which routing or fallback logic might hand a 5.2 conversation to another model. After that, rerunning evals across effort levels shows, in your own numbers, whether the bill actually falls 20% or whether extra thinking and subagents absorb the saving.
Sources
- What’s new in Claude Fable 5.2 - Claude Platform documentation (breaking changes, pricing, availability; checked November 11, 2026)
- Claude Fable 5.2 and Claude Mythos 5.2 migration guide - Claude Platform documentation (migration steps and checklist; checked November 11, 2026)
- Introducing Claude Fable 5.2 and Claude Mythos 5.2 - Anthropic announcement (November 9, 2026)
Was this article helpful?
Thank you!
Received. Thank you!