What came out on November 9
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2. The structure is a familiar one: same model, two levels of safeguards. Fable is the version available to everyone. Mythos is the same thing without some cybersecurity and biology filters, and it is reserved for organizations that pass Anthropic's verification programs, the Cyber Verification Program and the Life Sciences Verification Program.
The basics are unchanged from Fable 5.1: a one million token context window, 128,000 tokens of maximum output, adaptive thinking always on, five effort levels from low to max. On the API the model is called claude-fable-5-2 and it is available from day one on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry, as well as in the Claude apps on paid plans.
Three things change. The price drops by 20%. Output is up to 30% faster, and there is a fast mode that goes up to two and a half times faster, at double the price. The safety filters step in far less often.
The release lands a few days before Anthropic's stock market listing. That is not a colorful detail, and I'll come back to it below.
The price list goes down, the bill doesn't
The new price list: $8 per million input tokens and $40 output, against $10 and $50 for 5.1. Batch halves them, $4 and $20. Cache reads go from $0.25 to $0.20. Fast mode costs $16 and $80.
So far, the good news. The catch is that Opus 5.5 costs $4 and $20, exactly half. And the real bill does not depend on the price per token but on how many tokens the model burns to finish a task.
A table Anthropic included in the system card shows this well, on FrontierCode 2.0, an agentic coding benchmark. At low effort Fable 5.2 solves 34.2% of tasks at an average cost of $2.10 per task. At medium, 37.6% at $2.48. At high, 45.3% at $6.72. At xhigh, 49.1% at $14.19. At max, 45.7% at $44.70.
Two things stand out. Going from high to xhigh doubles the cost for four points. And max costs three times xhigh to do worse. The level you choose matters more than the price list.
External measurements confirm the picture. On Terminal-Bench 4.0 Vals AI measured $14.56 per test with Fable 5.2 against $12.04 with Opus 5.5. On CursorBench 4.0, at maximum effort, Cursor reports $32.79 against $13.43. The lower per-token price does not close the gap with Opus: it narrows it a little.
Where it really pulls ahead, and where it doesn't
Anthropic published a broad comparison, and it is worth reading it by splitting the benchmarks into two groups.
In the first group the lead over Opus 5.5 is a few points. FrontierCode 2.0, 49.1% against 45.5%. OSWorld 2.1, which measures computer use, 89.3% against 87.2%. FrontierSWE v2, 69.9% against 62.3%. It is a real improvement, but on coding and office automation Opus 5.5 stays very close at half the price.
In the second group the gap is wide. Terminal-Bench-Science, agentic scientific research: 74.6% against 58.7%. Humanity's Last Exam with tools: 74.2% against 67.7%. Chartography, reading charts without tools: 73.4% against 64.4%. And RoboDojo, a robotic control simulation, 64.9% against 39.4%, a test Anthropic itself notes it ran internally.
The most striking result is not in the table. A team of about a thousand Fable 5.2 agents formalized Fermat's Last Theorem in Lean, a language for machine-checked proofs, in 30 hours with minimal human intervention. The first formalization, done by an unreleased research model, took 11 days. Opus 5.5, with the same setup, was at 26% after 18 hours and then barely moving. To be clear: it is a formalization of a known proof, not new mathematics, and there is no external review yet.
The bottom line is that Fable 5.2 is worth the difference where the task is long, open-ended and hard. Where the task is well defined, the difference shrinks.
Get updates on Claude and AI for business
One email when there's something worth reading. No spam.
Evaluating Claude for your organization? Learn what it costs or which plan fits best
Fewer refusals, same boundaries
For people working in security or science, the most practical change is another one. According to Anthropic, Fable 5.2's filters block about 55% fewer requests than Fable 5.1: about 60% fewer in cybersecurity and 40% fewer in biology.
The boundaries do not move, though. The blocked categories are the same, and activities like exploit generation stay out of Fable. What improves is precision: fewer legitimate requests stopped by mistake.
When a refusal does come, the API flags it with an explicit reason and offers a fallback mechanism to another model. For Fable 5.2 the permitted fallback models are Opus 4.8 and Opus 5. If you have an application in production, it is worth configuring this rather than handling refusals as a generic error.
Mythos 5.2 is a different story. It requires verification, and for a European company the path is narrow today: I cover it in the analysis of the Mythos 5.2 system card.
Four things that break in your code
If you use Claude Code, claude.ai, Managed Agents or the Agent SDK, you don't have to do anything: Anthropic builds the requests. If your own code builds them, there are four points to check before changing the model name.
The first is about conversation history. On Fable 5.2 the API checks that nothing has changed before a thinking block: system prompt, tools, earlier messages. If something has changed, the request returns a 400 error, on any account. On Fable 5.1 the check only applied to accounts created from August 31, 2026, so code that worked on an older account can break when you switch models.
The second is about switching models mid-conversation. Only Fable 5.2 and Mythos 5.2 read Fable 5.2 thinking blocks. If a router moves a conversation to Opus 5.5, the accumulated reasoning is lost.
The third is computer use: the old computer_20251124 tool returns an error on every platform except Bedrock and has to be replaced with the new toolset.
The fourth is the advisor tool: a Fable 5.2 executor accepts only Fable 5.2 or Mythos 5.2 as its advisor.
Then there are behavior differences that break nothing but are noticeable. From medium effort up, the model thinks even when the prompt asks it to answer directly: for trivial requests the lever is lowering effort. At low effort, on long tasks, it sometimes stops before finishing. And it readily delegates to subagents, which can weigh heavily on what a session costs. The rest of the mechanics are in the Claude API integration guide.
Data, retention and Europe
One point to settle early in any compliance review. According to Anthropic's documentation, Fable 5.2 and Mythos 5.2 require 30-day data retention and are not available under zero data retention unless Anthropic expressly authorizes it. That is the price of safety monitoring on frontier models.
The alternative is Enterprise Frontier Safeguards, which move monitoring into the customer's infrastructure and are compatible with zero data retention. Today they are in private beta, with names like Bridgewater, Salesforce and Stripe. For a mid-sized European company they are not yet a concrete option.
On data residency the route is the usual one: go through one of the three clouds and pick the region, checking the model is there, because frontier models do not land everywhere on the same day. I wrote about how to approach this in an orderly way in Claude and GDPR for businesses.
What to do now
Anthropic says it in its own documentation: for most workloads you start with Opus, and move to Fable when your tests on Opus at high effort are not enough. It is honest advice and I agree with it.
In practice. If you use Opus 5.5 and you are happy with it, there is no reason to change everything. Take the two or three processes where Opus stops halfway or slips on long sequences, and try only those on Fable 5.2, starting at high effort, which is the API default. The full comparison, with a rule for deciding, is in Fable 5.2 or Opus 5.5.
If you are already on Fable 5.1, the switch is worth it: lower price and faster responses. But it is not a string change. Run the conversation history test, check your routers and computer use, and redo your tests across several effort levels. Anthropic has not yet announced a retirement date for 5.1, so there is time to do it properly.
If you haven't started yet, this release does not change your strategy. It changes the ceiling: the best model available today costs less than yesterday's.