Claude Fable 5.2 is cheaper and finally beats Opus 5.5 at coding, but the API change that will break integrations is preserved thinking

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, its top-tier models for "the most daunting technical challenges." They are the same underlying model: Fable ships with cybersecurity and biology safeguards and is open to every paid plan and the API (claude-fable-5-2), while Mythos runs without some of those safeguards and goes only to organizations accepted into Anthropic's Cyber and Life Sciences Verification Programs. Fable 5.2 costs $8 per million input tokens and $40 per million output, down from $10 and $50 for Fable 5.1, generates output up to 30% faster, and has an optional fast mode at double the price ($16/$80) for up to 2.5x the speed.

The announcement is long on showcase projects: 1,000 agents formalizing Fermat's Last Theorem in Lean in 30 hours, a prototype operating system written by more than 1,000 subagents over eight days, two solved open problems in black-hole physics, three circuit boards that worked without modification. Those are hard to check from the outside. The benchmark table, the footnotes and the API docs are easier, and they say a few things the headline doesn't.

Fable finally earns its price on coding

The most telling number in Anthropic's own table is the one for the previous model. On FrontierCode 2.0, Anthropic's chosen agentic-coding benchmark, Fable 5.1 scored 32.4%, 13 points below Opus 5.5's 45.5%. On that test the expensive tier was worse than the daily driver. Fable 5.2 reaches 49.1% at Xhigh effort, ahead of Opus 5.5, GPT-6 Astra (42.8%) and Gemini 4 Carbon (41.1%). At High effort it scores 45.3%, level with Opus 5.5's best, and Anthropic says it does that for roughly half the cost per task. On FrontierSWE v2 the jump is from 56.3% to 69.9%, against 62.3% for Opus 5.5.

The gains are not uniform, and the table shows where Fable 5.2 doesn't lead:

On AutomationBench, the business-workflow test, Gemini 4 Carbon is still ahead. On AA-Briefcase, Fable 5.2 wins on rubric score but Opus 5.5 keeps the higher Elo rating. The biggest single jumps are on the Remote Labor Index, which roughly doubles, and on RoboDojo, which Anthropic ran internally on its own harness. Anthropic's advice is consistent with that mixed picture: it still tells users to start with Opus 5.5 for "most coding and knowledge tasks" and move up to Fable only for ambitious, open-ended work.

"10.2% more vulnerabilities" is a ratio, and a footnote hedges it

The cybersecurity section says Mythos 5.2 found "10.2% more vulnerabilities than Claude Opus 5.5 and 14.4% more than Claude Mythos 5.1" on CVE Rediscovery, an internal test of re-finding 1,829 published CVEs in pre-fix snapshots of their projects. The table gives 89.6% for Mythos 5.2, 81.3% for Opus 5.5 (without safeguards) and 78.3% for Mythos 5.1. Those percentages are relative: 89.6/81.3 is 1.102 and 89.6/78.3 is 1.144. In absolute terms the gaps are 8.3 and 11.3 percentage points. Footnote 2 then adds that "on our own evaluations, vulnerability identification is roughly on par with Opus 5.5," pointing to the system card, and that the striking "overstated 1 in 20 findings" comparison comes from one tester's raw counts. A partner quote saying Mythos 5.2 "roughly doubled" Firefox vulnerability-discovery metrics is a testimonial, not something the evaluation section supports. Fable 5.2 users won't see most of this anyway: its classifiers block most security work and hand the request to Opus 4.8 (Opus 5 for biology), though Anthropic says false blocks are down about 55% from Fable 5.1.

A model deliberately held back on AI research

One paragraph deserves more attention than it will get. Anthropic says that during training it "took steps to limit advances in the specific research capabilities that could most directly accelerate AI development," so that on accelerator kernel optimization and improving LLM architectures, Fable 5.2 and Mythos 5.2 perform "at about the level of Opus 5.5." It says agentic coding, hardware work and ordinary ML research (including interpretability) were not affected, and that nothing is changed at inference time. Separately, a small set of topics in this domain, including kernel development on certain nonstandard accelerators, gets a transparent fallback to Opus 5. So a lab is now openly shaping training to keep one capability flat while the rest of the model improves. The claim that the cut is surgical rests on Anthropic's own evaluations; nobody outside has tested it yet.

The safety numbers are also more modest than the earlier Mythos history might suggest. In a dedicated test with more than 1,800 scenarios built around a deliberately weak sandbox, Mythos 5.2 used credentials or files planted outside it in 3.1% of scenarios, against 3.3% for Opus 5.5, 5.1% for Sonnet 5.5, and 39% and 21% for Mythos 5 and 5.1. Anthropic also reports that Fable 5.2 and Mythos 5.2 voice suspicion that they are being evaluated more often than any of its recent models, and says plainly that this "challenges our ability to assess how they will act in real-world settings."

The change developers will actually hit: preserved thinking, now strict

For API users the most consequential line was in a ClaudeDevs reply rather than the announcement: "To help prevent distillation, preserved thinking now covers all API accounts on Fable 5.2. Editing prior context can 400."

The preserved-thinking docs explain what that means. Each thinking block is bound to everything sent before it: the top-level system prompt, the tools list and the earlier messages. If any of that changes between requests, the block fails a prefix check. On Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5, accounts created before August 31, 2026 were let through: the stale thinking reached the model with no error. On Fable 5.2 and Mythos 5.2 there is no such grace, and every account gets a 400:

messages.1.content.0: Invalid `signature` in `thinking` block. The block is bound to a different conversation. Remove the block, or set `thinking.block_binding.prefix_mismatch_behavior` to "drop_block".

The list of things that count as an edit is long and full of ordinary harness habits: re-rendering the date or project instructions in the system prompt, shortening an old tool result, re-encoding an old image, adding a tool mid-session, an MCP server whose tool list changes between requests, a URL whose bytes change, rebuilding a saved session from templates the next day. The docs also warn that a clean run on your own key proves little, because users of your tool on other accounts or models can get the error when you don't. The escape hatch is the thinking-binding-controls-2026-08-01 beta header with prefix_mismatch_behavior: "drop_block", which drops the failing blocks and lets the request through, at the cost of the model re-deriving its reasoning and spending more output tokens doing it. The real fix is an append-only conversation, using mid-conversation system messages, tool_addition/tool_removal blocks and server-side compaction instead of rewriting history.

Two more restrictions arrive with 5.2. Fable 5.2 thinking blocks are tied to the account that produced them and are silently dropped if another, unlinked account sends them. And they are one-way: according to the docs' compatibility table, only Fable 5.2 and Mythos 5.2 can read Fable 5.2's thinking, so a router that moves a conversation from Fable 5.2 down to Opus 5.5 loses the earlier reasoning without an error. That fits the stated purpose. Anthropic says distillers operate through "thousands of fake Claude accounts" and that its countermeasures cut harvested data by 90% to 99%; binding reasoning to an account and a byte-exact history makes it much harder to replay. The cost lands on every developer whose agent loop rewrites context, and with Fable 5.2 there is no longer an old-account exemption to hide behind.

For subscribers, the launch also comes with usage-limit resets and a free Fable week on Pro, which we cover separately.