The essentials
Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026, positioning the pair as its most capable models for the hardest technical work: coding, finance, and the sciences. The two names hide no architectural secrets — Fable and Mythos are the same underlying model. Fable carries the cybersecurity and biology safeguards that make it available to all users, while Mythos strips some of those safeguards and is reserved for organizations accepted into Anthropic's expanded Life Sciences and Cyber Verification Programs.
For buyers, the practical change is speed at a lower price. Testers found the new models communicate more clearly than their predecessors, generate outputs up to 30% faster than Claude Fable 5.1 and Mythos 5.1, and are available at $8 per million input tokens and $40 per million output tokens.
Fast mode doubles throughput for a premium
Both models also ship with an optional fast mode that speeds up token generation by up to 2.5 times. It is priced at $16 per million input tokens and $80 per million output tokens, twice the standard rate.
| Access tier | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Standard | $8 | $40 |
| Fast mode — up to 2.5× faster generation | $16 | $80 |
Where the benchmarks moved
Anthropic's published benchmark table puts Fable 5.2 ahead of Opus 5.5 — its prior strongest daily-driver model — in almost every capability the company evaluated, and ahead of OpenAI's GPT-6 Astra and Google's Gemini 4 Carbon on most coding workloads. Scores reflect each model's highest-available effort with safeguards on unless noted.Anthropic announcement: Fable 5.2 and Mythos 5.2 evaluations
| Benchmark | Category | Fable 5.2 | Fable 5.1 | Opus 5.5 |
|---|---|---|---|---|
| FrontierCode 2.0 | Agentic coding | 49.1% (Xhigh) | 32.4% | 45.5% |
| FrontierSWE v2 | Agentic coding | 69.9% | 56.3% | 62.3% |
| AutomationBench | Business workflows | 51.4% | 31.4% | 42.5% |
| Humanity's Last Exam (with tools) | Multidisciplinary reasoning | 74.2% | 65.6% | 67.7% |
| OSWorld 2.1 (offline set, partial) | Computer use | 89.3% | 83.5% | 87.2% |
| RoboDojo Sim | Robotics | 64.9% | 27.6% | 39.4% |
Cost efficiency is the argument Anthropic makes around those scores: on FrontierCode 2.0, Fable 5.2 at High effort scored about level with Opus 5.5 at Max effort for roughly half the cost per task. On Humanity's Last Exam, Low-effort Fable 5.2 matched Opus 5.5's top score for less than one-tenth of the cost per attempt, and at Xhigh effort it reached 49.1% — the highest score Anthropic reports for any released model it has tested.
Coding results in production settings
Anthropic says more than 90% of the changes now merged into its own core codebase are written by Claude with human guidance. In one build, a team of agents running an earlier version of the model migrated a critical production service from Python to Rust — Anthropic reports the migrated service is more efficient, more reliable, and starts about 26 times faster.
The announcement's most-shared demonstration is an experiment in which one Anthropic engineer asked Fable 5.2 to design a prototype operating system. Starting from an empty folder, the model coordinated more than 1,000 subagents over eight days of round-the-clock work to produce a working OS that boots on a laptop — with its own core, drivers for disk, keyboard, and network, file storage, windows, a login screen, an installer, and built-in Claude functionality.
Projects we've been pushing along with other models for weeks got finished and approved by our teammates in a few days with Claude Fable 5.2, and the humans reviewing its code had less to say.
— Joe Haddad, Member of Technical Staff, Vercel
In one test combining country debt summaries, repayment history, and regional benchmarks, Claude Fable 5.2 generated SQL that returned the expected results. The SQL generated by two other models also ran successfully, but incorrectly summed country-level summary amounts, producing values 5–8x the expected amounts.
— Liyu Gong, Senior Manager of Machine Learning, Oracle
Early testers named in the announcement include Vercel, Shopify, Datadog, Unity, Uber, Oracle, iGent, Snowflake, and Cognition — with consistently reported strengths in completing long-horizon tasks without intervention and in exercising engineering judgment on ambiguous requests.
Science, hardware, and robotics beyond code
Two open problems in theoretical physics
Anthropic reports Fable 5.2 solved two open problems in black-hole physics, both connected to holographic duality. Working largely autonomously, the model completed a 2003 computation showing that black-hole formation stays abrupt in very small universes where only the quantum description applies, and it found the exact limit of a technique physicists have used for more than fifteen years to rule out candidate theories of quantum gravity. Anthropic published both results as preprints.
Protein folding with symmetry-aware merging
On gene-therapy delivery vehicles, Mythos 5.2 built a novel method for predicting the structure of a whole adeno-associated virus capsid — a protein shell made of dozens of individual proteins that tools like AlphaFold predict one at a time. The code ran existing open-source folding models on overlapping protein-chain groups and merged results using known capsid symmetries. In Anthropic's tests, the method accurately modeled 28 of the 34 capsids it attempted.
Formalizing Fermat's Last Theorem in 30 hours
A team of 1,000 Fable 5.2 agents fully formalized the proof of Fermat's Last Theorem in the Lean language in 30 hours, with minimal human intervention and no mathematical references beyond a list of intermediate theorem statements used for progress scoring. Anthropic's first formalization, by an unreleased research version of Claude, took 11 days with a purpose-built harness and human guidance. The Fable 5.2 run needed one-third as many action tokens and produced a proof seven times shorter; Opus 5.5 with the same setup completed only 26% of the formalization in 18 hours.
Hardware and robotics
Given a one-paragraph prompt, Fable 5.2 designed three working electronic devices — a home voice-assistant hub, a USB speakerphone, and a handheld virtual pet — each with its own four-layer circuit board, 3D-printed enclosure, and software, in collaboration with board manufacturer Diode. Every board was manufacturing-ready in under a day with minimal human input, and all three devices worked without electrical modifications. On the public RoboDojo benchmark for simulated two-arm robot control, the model scored 64.9% at Max effort — more than double GPT-6 Astra's 29.8% and Fable 5.1's 27.6%.
Safety posture and safeguards
Anthropic ran extensive pre-release safety testing, including novel evaluations probing specific concern areas such as sandbox escapes, and external testing by the UK AI Security Institute, Frontier Design, and METR. On the company's automated behavioral audit across roughly 1,900 scenarios, Mythos 5.2 — the same model without certain safeguards — scored 2.67 for misaligned behavior, better than Opus 5.5 (2.79), Sonnet 5.5 (3.10), and Mythos 5.1 (3.28), where lower scores are better.
In a dedicated adversarial sandbox-escape evaluation of more than 1,800 scenarios, the model used credentials or files outside its sandbox in 3.1% of cases — about as often as Opus 5.5 (3.3%), less often than Sonnet 5.5 (5.1%), and far below previous Mythos versions 5 and 5.1, which scored 39% and 21% respectively.
On safeguards for abuse: Fable 5.2 keeps robust cybersecurity and biology blocks with fewer false positives — Anthropic measured about 55% fewer unnecessary blocks than Fable 5.1 across claude.ai, Claude Code, and Cowork, split roughly as 60% fewer in cybersecurity and 40% fewer in biology. Blocked queries fall back to Opus 4.8 for cybersecurity and Opus 5 for biology. Anthropic also says Fable 5.2 carries its strongest-ever anti-distillation protections, which have reduced data harvested by illicit distillers by 90% to 99%, and that the preserved-thinking API feature now applies to all accounts.
One deliberate limitation is worth knowing for research teams: during training, Anthropic capped the model's advances in AI self-improvement capabilities — accelerator kernel optimization and improved LLM architectures — so Fable 5.2 and Mythos 5.2 perform at about Opus 5.5's level in those areas, with transparent fallbacks to Opus 5 for a small set of related topics.
Availability
Claude Fable 5.2 is available today on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure, with the model ID claude-fable-5-2 on the Claude Platform. Fast mode works on Claude Code, the Claude Platform, Claude Models on Google Cloud, and Microsoft Foundry. As a launch promotion this week, subscribers on eligible paid plans receive one to three Fable resets based on their subscription age, and Fable is included in up to 50% of weekly usage limits for Pro plans and standard Team-plan seats.
Access to Mythos 5.2 — and to its fast mode — is limited to vetted organizations through the Cyber Verification Program and Life Sciences Verification Program. Claude Enterprise customers can also use Claude Security, which now includes Mythos 5.2 among the models scanning codebases for vulnerabilities.
Anthropic's own product guidance keeps expectations calibrated: for most coding and knowledge work, start with Claude Opus 5.5 and escalate to Fable or Mythos only for ambitious, open-ended technical problems — or use Fable 5.2 as an Advisor that other models consult, or as an Orchestrator that delegates simpler tasks to smaller models.
Why this story is trending
- The result that traveled fastest was the 30-hour, 1,000-agent formalization of Fermat's Last Theorem — Anthropic's second formalization of the proof, cutting 11 days to 30 hours.
- Two verified open problems in black-hole physics solved autonomously, plus a working prototype operating system built by coordinating more than 1,000 subagents over eight days, gave the launch unusually concrete demonstrations.
- The pricing move is aggressive for a frontier-flagship tier: $8/$40 per million tokens with a 2.5× fast mode, alongside claims of roughly half Opus 5.5's cost per coding task.
- Security teams are paying attention because Claude Mythos 5.2 adds more vulnerability-hunting capability to the restricted Verification Program track — one security company found one in six of the model's confirmed bugs was missed by all other frontier models they tested.
- The release lands in the middle of the frontier-race benchmark cycle, directly against GPT-6 Astra and Gemini 4 Carbon on agentic coding, robotics, and computer-use scores.
What to watch next
- Independent replication of the headline demonstrations — the physics preprints, the capsid-folding method, and the Lean formalization are all worth following through Anthropic's published artifacts.
- How quickly real-world coding teams standardize on efficient-tier routing: Opus 5.5 as the daily driver, Fable 5.2 as the escalation path, per Anthropic's guidance.
- Fast-mode economics in practice — whether the 2.5× throughput premium at $16/$80 pencils out for high-volume agent workloads.
- Mythos access expansion through the Cyber and Life Sciences Verification Programs, and what its adoption in Claude Security means for enterprise vulnerability scanning.
- The system card — Anthropic's published Claude Fable 5.2 and Mythos 5.2 System Card — for methodology behind the headline benchmark and safety numbers.
Further reading
Building a product in this space?
MORE FROM THE NEWSROOM
Related news
Artificial IntelligenceUS Moves to Regulate Remote Access to Advanced AI Chips Through Overseas Data Centers
The US Department of Commerce has released an interim final rule targeting remote access to advanced American semiconductors through overseas data centers. The measure introduces licensing requirements for certain high-volume computing arrangements serving companies headquartered in restricted destinations and adds destination-based controls for advanced chip exports to five countries.
Artificial IntelligenceOpenAI Launches ChatGPT Wallet, Letting AI Agents Complete Purchases With User Approval
OpenAI has introduced ChatGPT Wallet, a payment feature that allows its AI agents, called Dots, to complete online purchases after users review and approve each transaction. Developed with American Express, Mastercard and Visa, the initial rollout is limited to eligible US ChatGPT Pro subscribers.
Artificial IntelligenceCerebras Raises 2026 Revenue Outlook After 55% Quarterly Growth, but GAAP Losses Widen
AI hardware company Cerebras reported record third-quarter revenue of $210.8 million and raised its full-year 2026 revenue outlook. Its cloud and inference business grew rapidly, but the results also revealed a much larger GAAP net loss, highlighting the financial challenges of scaling advanced AI infrastructure.
