Claude Fable 5.2 and Mythos 5.2: Frontier Release
Quick verdict
Claude Fable 5.2 and Mythos 5.2 arrived November 9, 2026 at $8/$40 per million tokens — 20% below Fable 5.1 — with outputs up to 30% faster and an optional 2.5x fast mode at $16/$80. Fable is the safeguarded public version; Mythos is the same model with fewer blocks, gated behind Anthropic verification programs. Fable 5.2 tops Anthropic-reported FrontierCode 2.0 at 49.1% and more than doubles Fable 5.1 on simulated robotics.
Official sites
Check current product details and plans with each provider.
What Shipped
On November 9, 2026, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2. They are the same underlying model. The difference is safeguards:
- Fable 5.2 carries strong cybersecurity and biology safeguards that block fewer legitimate requests than 5.1, which is what allows Anthropic to offer it to everyone.
- Mythos 5.2 is the same model with permissive capability, available only to organizations approved through Anthropic's Cyber Verification Program or Life Sciences Verification Program.
That split is the most interesting design decision in the release. Anthropic is selling capability and constraint separately instead of withholding the model.
Pricing and Speed
| Item | Fable 5.2 | Fable 5.1 | Opus 5.5 |
|---|---|---|---|
| Input / 1M | $8 | $10 | $4 |
| Output / 1M | $40 | $50 | $20 |
| Cache read / 1M | $0.20 | $0.25 | $0.20 |
| Cache write / 1M | $10 | — | $5 |
| Fast mode | $16 / $80 | — | available at 2x |
Two things stand out. First, Fable 5.2 is 20% cheaper than Fable 5.1 at the same tier — price went down while capability went up. Second, outputs are up to 30% faster, and the optional fast mode adds up to 2.5x generation speed for double the per-token rate.
Anthropic's launch promotion gives eligible paid subscribers one to three Fable resets and includes Fable for up to 50% of weekly usage limits on Pro plans and standard Team seats.
Benchmarks
Anthropic's published comparison table (vendor-reported on its own harness):
| Benchmark | Fable 5.2 | Fable 5.1 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|---|
| FrontierCode 2.0 (agentic coding) | 49.1% (Xhigh) | 32.4% | 45.5% | 42.8% | 41.1% |
| FrontierSWE v2 | 69.9% | 56.3% | 62.3% | 65.5% | 61.2% |
| AutomationBench | 51.4% | 31.4% | 42.5% | 41.4% | 53.9% |
| Terminal-Bench-Science 0.1 | 74.6% | 75.3% (Mythos 5.2) | 52.6% | 58.7% | 68.1% |
| CyScenarioBench † | 76.3% (Mythos 5.2) | 61.7% (Mythos 5.1) | 67.6% (no safeguards) | 61% (GPT-6 Cyber) | — |
| CVE Rediscovery | 89.6% (Mythos 5.2) | 78.3% (Mythos 5.1) | 81.3% (no safeguards) | — | — |
| De novo binder design (24 hr) | 85.1% (Mythos 5.2) | 79.1% (Mythos 5.1) | 82.6% (no safeguards) | — | — |
| WANDR (agentic search) | 82.1% | 68.7% | 72.3% | — | — |
| OSWorld 2.1 (computer use) | 89.3% | 83.5% | 87.2% | 79.5% | 81.5% |
| RoboDojo Sim ‡ | 64.9% | 27.6% | 39.4% | 29.8% | — |
| AA-Briefcase v1.1 | 66.1% | 61.5% | 61.2% | 52.0% | 61.8% |
| Humanity's Last Exam (with tools) | 74.2% | 65.6% | 67.7% | 57.2% | — |
| Chartography (no tools) | 73.4% | 44.8% | 64.4% | 71.0% | 72.5% |
Read that table with two caveats. Anthropic ran the competitor models itself on several of these harnesses, so the comparison is not independent. And some rows use safeguards-off configurations for Opus 5.5 and Astra, which are noted in the footnotes but easy to miss.
The two rows that survive scrutiny best are the ones where the gap is implausibly large to be harness noise:
- RoboDojo Sim: 64.9% versus 27.6% for Fable 5.1. More than double on a simulated two-arm manipulation task, with no explicit robotics training.
- FrontierSWE v2: 69.9% versus 56.3%. Thirteen points on software engineering, which is the area Anthropic's customers actually buy.
The weak row is Terminal-Bench-Science at 74.6%, essentially flat against Mythos 5.2's 75.3% — Anthropic labels Mythos 5.2 there, which suggests the scientific-research ceiling did not move.
Cost Per Task Is The Real Story
Anthropic's framing is that raw scores matter less than score-per-dollar. Two claims worth taking seriously:
- At High effort, Fable 5.2 scores about the same as Opus 5.5's best on FrontierCode 2.0 (45.3% vs 45.5%) for roughly half the cost per task.
- At Medium effort — the default in Claude apps and Claude Code — Fable 5.2 solves challenging tasks with fewer tokens than Fable 5.1, and on WANDR it matches Opus 5.5's top score at about 40% of the cost.
If those hold on your own workload, the interesting move is not "upgrade to Fable" but "run Fable at medium effort instead of Opus at max." That is a routing decision, and it depends entirely on your task mix.
What It Actually Did
Anthropic lists concrete prerelease work, which is more useful than benchmark deltas:
- Software migrations its own engineers had considered infeasible
- Three working electronic devices, designed from scratch, that needed no electrical modifications
- Two open problems in black-hole physics — notably completing a 2003 calculation about whether small universes also undergo an abrupt black-hole transition, which had been "too laborious to complete"
- Robotics, including solving a puzzle cube, unscrewing a bottle cap, and drawing a robot
- Fermat's Last Theorem, fully formalized in 30 hours by a team of 1,000 coordinated Fable 5.2 agents — versus 11 days for an earlier unreleased research version with human guidance
The Fermat result is the most striking and the most easily misread. Formalizing a known proof in a proof assistant is not the same as discovering a proof. The 1,000-agent coordination result is the real claim: the model organized its own division of labor with minimal human intervention.
Third-Party Signals
Anthropic quotes external testers, and a few are worth noting because they describe failure modes rather than praise:
- Mozilla: Mythos 5.2 "roughly doubled the vulnerability discovery metrics of Mythos 5.1" across successive Firefox scanning campaigns, with each round leaving "fewer, subtler bugs to find."
- iGent: built a hypervisor whose memory isolation is backed by more than 3,500 machine-checked proofs in a single session.
- Oracle: in a SQL generation test, two other models produced queries that ran successfully but summed country-level amounts 5–8× too high; Fable 5.2 returned the expected results.
- Snowflake: strongest measured model on its 103-task data-engineering benchmark.
- Cognition: "says it doesn't know things instead of hallucinating" — it retracted an estimate rather than guess.
Anthropic's own footnotes add an important correction: one tester's raw CVE counts were much higher for Mythos 5.2, but on Anthropic's own evaluations, vulnerability identification is roughly on par with Opus 5.5.
Safety and the Monitoring Tradeoff
Anthropic says it ran extensive prerelease testing, including new evaluations targeting sandbox escapes, and that high-stakes misalignment risk remains low with performance "as well as, or better than" recent models on most alignment measures.
Two details deserve more attention than the summary paragraph:
Safeguards block less. Fable 5.2's cyber and biology safeguards are described as equally robust as 5.1's but producing fewer unnecessary blocks. That is the change that makes a public Fable release possible — and it means "no block" is a weaker signal of safety than it was.
Monitoring looks at usage windows. Anthropic states that models like Fable 5.2 and Mythos 5.2 rely on safeguards that monitor a window of usage data rather than reviewing each request, because some misuse is only visible across a pattern. Enterprise Frontier Safeguards customers can run zero data retention; other customers have 30-day retention for safety monitoring.
Also new: Anthropic calls this its strongest-ever defense against distillation, claiming the techniques reduce data harvested by distillers by 90% to 99%, and preserved thinking now applies to all accounts so distillers cannot edit Claude's prior context to extract reasoning.
Availability
- Fable 5.2 is available on all platforms, including AWS, Google Cloud, and Microsoft Azure. API model id:
claude-fable-5-2. - Mythos 5.2 requires approval through the Cyber Verification or Life Sciences Verification Program.
- Fast mode is available on Claude Code, the Claude Platform, Claude on Google Cloud, and Microsoft Foundry. Mythos fast mode is restricted to verification-program organizations.
When to Use It
Anthropic's own guidance is worth repeating because it argues against the obvious move: start with Opus 5.5 for most coding and knowledge work, then move up to Fable for the hardest open-ended problems or down to Sonnet 5.5 or Haiku 5.5 for volume.
Fable 5.2 can also act as an Advisor (smaller models consult it on hard subproblems) or an Orchestrator (it delegates simple work to cheaper models). For teams already routing across tiers, that is the more interesting use than replacing the daily driver.
Limitations
- Vendor benchmarks. Anthropic ran competitor models on its own harnesses for several rows.
- Safeguards-off comparisons. Some Opus 5.5 and Astra figures in the table are without safeguards, which is not how customers run them.
- Formalization is not discovery. The Fermat result validates coordination and proof engineering, not mathematical insight.
- Science ceiling flat. Terminal-Bench-Science did not move versus Mythos 5.2.
- Fast mode is expensive. Doubling the price changes the cost-per-task math.
The Verdict
Fable 5.2 is a genuine generational step for the hardest technical work: cheaper than its predecessor, faster, and measurably better at software engineering and physical-world control. The Fermat formalization and the hypervisor proofs are the kind of result that changes what a small team can attempt.
The release is equally notable for how Anthropic packaged it. Splitting capability from safeguard level — Fable for everyone, Mythos behind verification — is a more honest structure than withholding the model, and the "fewer unnecessary blocks" improvement is what makes it work.
For most teams the practical decision is not "switch to Fable." It is "run Fable at medium effort on the hard problems and keep Opus 5.5 as the default." Measure cost per finished task before changing anything else.
Official links
Related Articles
Frequently Asked Questions
What is Claude Fable 5.2?+
What is the difference between Fable 5.2 and Mythos 5.2?+
How much does Claude Fable 5.2 cost?+
How much better is Fable 5.2 than Fable 5.1?+
What did Claude Fable 5.2 actually accomplish in testing?+
Should I switch from Opus 5.5 to Fable 5.2?+
Pros
- 20% cheaper than Fable 5.1 at $8/$40 per million tokens, with cache reads at $0.20
- Outputs up to 30% faster, plus an optional 2.5x fast mode at $16/$80
- FrontierCode 2.0 at 49.1% — the highest Anthropic has measured, ahead of GPT-6 Astra and Gemini 4 Carbon
- RoboDojo Sim 64.9%, more than double both Fable 5.1 and GPT-6 Astra on the same harness
- Safeguards are as strong as Fable 5.1 but block less, which makes Fable usable by everyone
Cons
- Mythos 5.2, the version with fewer blocks, requires verification-program approval
- Most benchmark numbers are Anthropic-reported, including competitor runs on its own harness
- Anthropic still recommends Opus 5.5 as the daily default — Fable is for the hardest work
- Fast mode doubles the per-token price
- Frontier safeguard monitoring uses a window of usage data, not per-request review
Keep reading