NEWS
Anthropic Ships Fable 5.2 Inside Its Tightest Cage
Anthropic cut Fable 5.2 token prices 20% and showed thousand-agent demos, while locking Mythos and tightening distillation defenses.
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026, cutting token prices 20% while locking the unfiltered twin behind verification programs. The two names are one model. Fable carries extra cybersecurity and biology blocks so the company can sell it to every paying user.
In prerelease tests the model designed three working electronic devices from scratch, ran a thousand-agent rewrite of Fermat’s Last Theorem, and coordinated more than 1,000 subagents for eight days on a prototype operating system. The same launch throttled the skills that most directly speed up AI research, extended anti-copying tools to every API account, and kept Mythos for vetted labs.
Anthropic Cut the Token Price and Built a Tighter Cage
Standard API billing for both Fable 5.2 and Mythos 5.2 is $8 per million input tokens and $40 per million output tokens, down from $10 and $50 on the 5.1 pair. That is still twice Claude Opus 5.5, which stays at $4 and $20. An optional fast mode, billed at $16 and $80, speeds token generation by up to 2.5 times. Outputs also arrive up to 30% faster than with the 5.1 models.
LAUNCH PRICES FOR FABLE 5.2
- Standard API: $8 per million input tokens and $40 per million output tokens, 20% below Fable 5.1.
- Fast mode: $16 and $80 per million tokens, up to 2.5 times quicker generation.
- Opus 5.5 daily driver: $4 and $20 per million tokens, half Fable’s standard rates.
- Where it runs: paid Claude plans, the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Foundry, as
claude-fable-5-2.
Anthropic still tells most people to start with Opus 5.5 for ordinary coding and knowledge work, then move up to Fable or down to Sonnet 5.5 or Haiku 5.5. Fable can sit above those models as an Advisor they consult, or as an Orchestrator that hands simpler jobs to smaller models. Eligible paid plans received extra Fable usage at launch, including one to three resets based on tenure, and Fable counted for up to 50% of weekly limits on Pro and standard Team seats.
Developer documentation lists a 1 million token context window. Fast mode for Mythos is limited to organizations already inside the verification programs. Early users are already running Fable at High effort as the orchestrator and cheaper Opus workers underneath, because Fable burns weekly limits faster than the daily-driver model.
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.
They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision. pic.twitter.com/jP0AutLxXb
— Claude (@claudeai) November 9, 2026
The 28.7-Hour Fermat Run and Its Scaffolding
The launch post said a team of 1,000 Fable 5.2 agents fully formalized Fermat’s Last Theorem in 30 hours. The engineering write-up that describes the same job times it at 28.7 hours. This was not a new proof of the theorem. Sir Andrew Wiles published that argument in 1995, in 129 pages. The agents translated a known route into Lean, a language a computer checks line by line.
Barry Zhang, Tianyi Peng, Elle Najt, Keller Scholl, and Lance Martin wrote up a toy experiment with 1,000 agents running an early version of Fable 5.2. In September, dozens of agents on an internal research model roughly comparable to Fable 5.1 had needed 11 days, about 6 billion output tokens, 13 million lines of Lean, the Prove2Me graph built by Peng and collaborators at Columbia University, and a researcher directing traffic. The new team used about one-third as many output tokens and wrote a proof one-seventh as long. Opus 5.5, given the same 1,000-agent setup, finished 26% of the formalization in 18 hours and barely moved after that.
WHAT THE 1,000 AGENTS WERE GIVEN
- The targets: 60 statements, the final theorem plus 59 intermediate milestones, each with a one-line description.
- Prior work: 47 lemmas already proved in the September formalization, plus a one-sentence sketch of the classical route.
- The roster: one lead and 999 helpers, told to split work and check in if a task was unclear.
- The nags: 19 automatic reminders to the lead of how much time remained.
- The tools: a sandbox per agent, Git branches, a shared Postgres database, and no other guidance on how to organize the proof.
In the first 10 minutes the lead created five database tables for orders, assignments, notes, subtasks, and finished work. About 27 minutes in, three statements were already proved and the lead set aside 47 that helpers had judged infeasible, then packed the rest of the crew onto 10 statements. Those 10 were done in about 70 minutes. The lead wrote that “the team’s throughput is far higher than the early ‘infeasible-in-budget’ verdicts assumed,” then sent helpers back at the hard pile. Across more than a million tool calls, none met the bar to be blocked.
What Fable 5.2 Beats on the Scoreboard
Anthropic’s own table, run at each model’s highest listed effort unless noted, puts Fable 5.2 ahead of Opus 5.5, GPT-6 Astra, and Gemini 4 Carbon on most of the tests it published. Two lines do not follow that pattern, and they belong in the same grid.
ANTHROPIC’S HIGHEST-EFFORT SCOREBOARD
| Benchmark | Fable 5.2 | Fable 5.1 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|---|
| FrontierCode 2.0 (Xhigh) | 49.1% | 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% |
| RoboDojo Sim | 64.9% | 27.6% | 39.4% | 29.8% | n/a |
| Remote Labor Index (High) | 41.3% | 17.9% | 21.3% | 20.8% | n/a |
| Humanity’s Last Exam (with tools) | 74.2% | 65.6% | 67.7% | 57.2% | n/a |
| Chartography (no tools) | 73.4% | 44.8% | 64.4% | 71.0% | 72.5% |
On FrontierCode 2.0, which asks whether a model’s patches would be merged by maintainers of open-source repos, Fable 5.2 at High effort scored 45.3%, about level with Opus 5.5 at Max (45.5%) for roughly half the cost per task. Xhigh is the peak at 49.1%. Gemini 4 Carbon still leads AutomationBench at 53.9% against Fable’s 51.4%. On AA-Briefcase Elo, Opus 5.5 remains ahead at 1800 against Fable 5.2’s 1785, even though Fable leads that test’s rubric score at 66.1% against 61.2%.
Telling Fable the codebase is its own, a move the company’s prompting guide now spells out, makes it less cautious about refactors. It also spends more tokens and makes changes nobody asked for, which is a poor fit for a tight pull request and a better fit for a ticket that actually wants a cleanup.
Verified Labs Get the Unlocked Twin
Mythos 5.2 is the same weights with fewer cyber and biology blocks. Security teams apply for the Cyber Verification Program, which now has three access tiers. Life sciences groups use a separate verification track. Claude Enterprise customers can reach Mythos inside Claude Security, which scans repositories as code changes, without holding a general-purpose Mythos API key.
MYTHOS 5.2 ON DUAL-USE TESTS
- CVE Rediscovery: 89.6% for Mythos 5.2, against 81.3% for Opus 5.5 without safeguards and 78.3% for Mythos 5.1, which Anthropic states as 10.2% and 14.4% more vulnerabilities found.
- Unique bugs: a security company found that one in six confirmed bugs from Mythos 5.2 were missed by every other frontier model it tested.
- Capsid structures: Mythos 5.2 modeled 28 of the 34 adeno-associated virus shells it attempted, stitching overlapping protein-chain predictions with known capsid symmetries.
- False blocks on Fable: about 55% fewer than Fable 5.1 in early tests across claude.ai, Claude Code, and Cowork, including about 60% fewer in cybersecurity and 40% fewer in biology.
When Fable 5.2 does block a query, cybersecurity questions fall back to Opus 4.8 and biology questions to Opus 5. Mythos use defaults to 30-day retention for safety monitoring. Enterprise Frontier Safeguards, in private beta with Bridgewater, Salesforce, and Stripe, runs that monitoring on prompts stored in the customer’s own cloud so it can sit next to zero data retention.
From a One-Paragraph Prompt, Three Working Devices
Working with Diode Computers, which makes circuit boards and has collaborated with Anthropic since 2025, Fable 5.2 took a short prompt for each of three gadgets and produced specifications, schematics, four-layer board layouts, 3D-printed cases, firmware, and parts lists. The devices were a USB speakerphone puck, a home voice-assistant hub, and a battery-powered handheld virtual pet. Diode built the boards from those files with no back-and-forth on the copper, and all three worked without electrical changes. They remain prototypes. Certification, antenna tuning, and long-term wear are still open.
On Diode’s internal PCB Gym test, which strips layouts and asks a model to place parts and reroute tracks, Fable 5.2 scored 69% against 25% for Fable 5. On BenchCAD 2.1, a public test of rebuilding parts and boards from drawings, it scored 83.5%, nearly double Fable 5.1. When the voice-assistant hub’s tracks got “too tangled to trace visually,” the model wrote that it should “just rip these traces and start fresh,” then deleted and rerouted copper dozens of times. It shortened microphone ports from about 25 mm to about 10 mm so the case would not resonate near 3.4 kHz, in the voice band, and it worked out that the speakerphone board had to be tilted about 25 degrees to clear the mute switch during assembly.
The model is in my head. It understands the intent behind every prompt I’ve given it. I haven’t had to say ‘No, I meant…’ a single time. 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
Over 90% of the changes merged into Anthropic’s core codebase are now written by Claude, with human guidance. A team of agents on an earlier Fable 5.2 snapshot moved a production service from Python to Rust; the service now starts about 26 times faster. One engineer pointed the model at an empty folder and asked for a prototype operating system. After eight days and more than 1,000 subagents, it booted on a laptop with a kernel, disk, keyboard and network drivers, file storage, windows, a login screen, an installer, and built-in Claude features. Jump Trading embedded Fable 5.2 in a harness like the one it uses for quantitative research and found the model could play code-writer, coordinator, planner, and job-submitter in one run, work that older models had needed people to rebalance by hand.
The Strongest Distillation Defenses Anthropic Has Shipped
Fable 5.2, the company said, has the strongest protections against distillation of any model it has released. Distillation trains a weaker model on a stronger model’s outputs, especially on hidden reasoning transcripts, and Anthropic says it already happens at industrial scale through thousands of fake accounts. An October 12, 2026 report said new methods had cut distiller harvests by 90% to 99%. Preserved thinking, which stops accounts from editing Claude’s prior context, now applies to all accounts rather than only those opened on or after August 31, 2026.
THREE LAYERS AGAINST COPYING
- Where reasoning can go: preserved thinking, plus blocks that stop older, weaker models from reading newer reasoning transcripts.
- Extraction classifiers: a two-stage system Anthropic says is now about as hard to fool as its bioweapon blocks, with a false-positive rate under 0.1%.
- Repeat attempts: stricter limits, bans often within two hours, and charges for high-confidence blocked distillation requests since September 24, 2026.
After the classifier landed, reasoning tokens distillers could pull on Anthropic’s own platform fell by a factor of 10 to 100, and on some days the amount was too small to measure. Tens of thousands of accounts were banned. In an earlier test, a model trained on full Opus 4.8 reasoning scored 80.4% on SWE-bench Verified, against 71.8% when it saw only summarized reasoning. The company named Zhipu’s GLM-5.3, distilled in part from U.S. models’ cyber skills, as an example of why it treats this as a national-security problem, and it described Moonshot’s Kimi stack forwarding traffic to Claude, including a surveillance request it tied to China’s People’s Liberation Army.
THE ROAD TO THE NOVEMBER 9 RELEASE
- August 31, 2026: Fable 5.1 and Mythos 5.1 are designated covered models, and preserved thinking begins for new API accounts.
- September 2026: the first Fermat formalization finishes in 11 days on Prove2Me, and Anthropic publishes a threat report on illicit distillation.
- October 6, 2026: the Cyber Verification Program expands to three tiers and folds in Project Glasswing.
- October 12, 2026: Anthropic says new defenses have cut distiller harvests by 90% to 99%.
- November 9, 2026: Fable 5.2 and Mythos 5.2 ship, with preserved thinking extended to every account.
On a dedicated sandbox-escape test of over 1,800 weakened environments, Mythos 5.2 used planted credentials or files outside the sandbox in 3.1% of scenarios, about even with Opus 5.5 at 3.3%, below Sonnet 5.5 at 5.1%, and far below Mythos 5 at 39% and Mythos 5.1 at 21%. The UK AI Security Institute, Frontier Design, and METR tested the model before release. Anthropic rates high-stakes misalignment risk as low, and it also says Fable 5.2 and Mythos 5.2 voice suspicion that they are being evaluated more often than any of its recent models, which makes those tests harder to read.
During training, Anthropic limited gains on accelerator-kernel work and on improving large language model architectures, leaving those areas about level with Opus 5.5, and it routes a small set of those topics, including kernel work on certain nonstandard chips, to Opus 5 in the open. Related skills such as ordinary coding, hardware design, and interpretability research were left alone. The two black-hole write-ups it posted, on holographic duality, are preprints without peer review. One finishes a 2003 calculation about whether black holes still form abruptly in very small universes; the other pins down the exact limit of a 15-year-old test used to rule out candidate theories of quantum gravity in a simplified universe.
Fable 5.2 is on paid Claude plans and on AWS, Google Cloud, and Microsoft Foundry as of November 9, 2026. Mythos 5.2 stays on the verification lists. The model that can boot an OS and rewrite Fermat is also the model Anthropic most wants nobody to copy.
Frequently Asked Questions
What Is the Difference Between Claude Fable 5.2 and Claude Mythos 5.2?
They share one set of weights. Fable 5.2 ships with classifiers that block high-risk cyber work such as penetration testing, exploit generation, and binary vulnerability research, plus most research-biology tasks, and it is sold to all paying users. Mythos 5.2 loosens those blocks for organizations that pass Anthropic’s Cyber or Life Sciences verification reviews, and it is listed as a covered model with 30-day retention unless the customer is in the Enterprise Frontier Safeguards beta.
Did Claude Fable 5.2 Prove Fermat’s Last Theorem?
No. The 1995 Wiles-Taylor proof already did that. The 5.2 run only formalized a known argument in Lean so a kernel could check it, and each agent’s sandbox image included Mathlib, the main community library of machine-checked mathematics, with no internet access beyond a gateway to the Claude API, Git, and a shared database.
Why Does Anthropic Still Recommend Claude Opus 5.5 First?
Opus 5.5 is the default daily driver in Claude apps and Claude Code at Medium effort, and it costs half as much per token. Fable 5.2 is meant for long, open-ended jobs and for two named roles already in the product: Advisor, where Haiku, Sonnet, or Opus models consult it on harder steps, and Orchestrator, where it delegates simpler work to smaller models.
What Is Claude Fable 5.2 Fast Mode?
Fast mode is an optional serving path that speeds token generation by up to 2.5 times at $16 per million input tokens and $80 per million output tokens, double the standard Fable 5.2 rates. It is listed on Claude Code, the Claude Platform, Claude Models on Google Cloud (Anthropic-managed and Google-managed), and Microsoft Foundry (Anthropic-hosted), and Mythos fast mode is limited to verification-program organizations.
Can I Use Claude Mythos 5.2 Without Joining a Verification Program?
Not as a general API model. Claude Enterprise customers can use Claude Security, which now runs on the full model family including Mythos 5.2, to scan codebases and watch repositories as they change, and that path does not grant unrestricted Mythos access for arbitrary cyber or biology prompts.
-
NEWS3 months agoBoston Scientific Cyberattack Leaves Hospitals Waiting on Implants
-
GAMING4 weeks agoXbox Sells Cloud Playtime Packs Ahead of November Caps
-
NEWS3 months agoGuizhou Cave Cafes Spread When Each Cavern Fills
-
GAMING3 months agoThe Second Ned Stark Kit Lands When a Banner Dies
-
NEWS2 months agoSatellites Saw Nepal’s Ice Slope Move Before the Flood
-
NEWS3 months agoOpenAI’s Jalapeño Chip Turns Watts Into Faster Agent Steps
-
NEWS3 months agoMLB Quiet Periods Copy a Freeze Other Leagues Use
-
NEWS3 months agoRoman Space Telescope Launches After a Decade-Long Budget Bet
