Is Google back?11 minGemini 4 Carbon topped the leaderboards, then lost the crown the day before most customers could touch it. Meanwhile Google's share of enterprise AI spend fell from 21% to 9%. Here's the three-part test for whether Google is actually back, or just back on the leaderboard.Read now →

A Thousand Claude Agents Formalized Fermat's Last Theorem in 30 Hours. That Is Not Even the Wildest Part
Anthropic's Claude Fable 5.2 tops the leaderboards, built a bootable operating system with 1,000 subagents and controlled real robot arms. It also costs more per task than almost anything else. Here is what the benchmarks, the price war and the safety caps actually mean.
Contents
- The 30-hour proof
- Two names, one model
- The naming mystery
- Where the benchmark gains actually live
- The over-deliberation problem
- Cheaper per token, pricier per task
- Long-horizon work: an OS from an empty folder
- Physics, proteins and circuit boards
- The robotics bet
- The safety story is the strangest part
- The IPO backdrop
- Why this matters
- How to actually use it
- Risks and counterpoints
- FAQ
- The bottom line
The 30-hour proof
Here is the headline that made people stop scrolling.
Anthropic gave roughly 1,000 Claude Fable 5.2 agents a bare-bones setup and asked them to formalize the proof of Fermat's Last Theorem in Lean, the programming language mathematicians use to make a computer check every step of a proof.
They finished in about 30 hours.
Anthropic's previous attempt, run by an unreleased research model with a purpose-built harness and a human researcher guiding it, took 11 days. This time, the company says, the agents invented their own ways to coordinate, used one-third as many action tokens, and produced a proof seven times shorter.
Then came the comparison chart. Anthropic ran the same 1,000-agent setup with Opus 5.5, its current everyday flagship. At the 18-hour mark, Opus had completed about 26% of the formalization. Fable was far ahead. After 30 hours, Fable was done and Opus had barely moved.
That one chart explains the whole release. Fable 5.2 is not mainly a better chatbot. It is a better long-distance runner.
The fine print you should know
Before anyone calls this a new mathematical breakthrough: it is not. Andrew Wiles proved the theorem in 1995. Formalization means translating an existing proof into code a computer can verify.
Reporting from Tech Times also points out that the two runs did not start from the same place. Anthropic's engineering write-up says the agents got 60 target statements, a one-sentence sketch of the classical route, the needed definitions and 47 lemmas already proved in the earlier formalization. And at launch, there was no independent review of the new proof comparable to the one the September result received.
So: real, impressive, and worth reading carefully.
Two names, one model
Anthropic launched two products on November 9, 2026: Claude Fable 5.2 and Claude Mythos 5.2.
They are the same underlying model with different guardrails.
- Fable 5.2 is the public version. It ships with strong cybersecurity and biology safeguards that block things like exploit generation and most research-biology tasks.
- Mythos 5.2 drops those restrictions and is available only to vetted organizations through Anthropic's Life Sciences and Cyber Verification Programs.
The guardrails got less annoying, too. Anthropic says Fable 5.2 produced about 55% fewer unnecessary blocks than Fable 5.1 in early testing, roughly 60% fewer in cybersecurity and 40% fewer in biology. When a request is blocked, the answer comes from an older model instead (Opus 4.8 for cyber, Opus 5 for biology).
Fable is live on all major platforms, including AWS, Google Cloud and Microsoft Azure. For launch week, Anthropic is giving eligible paid subscribers one to three extra "Fable resets" depending on how long they have subscribed. Pro plans can spend up to 50% of their weekly limits on Fable.
The naming mystery
One small oddity: Opus is already on version 5.5, while this flagship launches as 5.2.
Plenty of people expected "Fable 5.5." Anthropic has not explained the numbering. Fable and Mythos simply follow their own track: 5.1 shipped in September, and 5.2 follows it.
It is a reminder that version numbers across model families now mean very little. Compare benchmarks and prices, not decimals.
Where the benchmark gains actually live
On the independent Artificial Analysis Intelligence Index, Fable 5.2 at max effort scores 65, ranking first of the 74 models listed. Launch-week coverage put Opus 5.5 several points behind and Fable 5.1 well behind that.
But the more useful story is in the shape of Anthropic's own benchmark table. The jumps are huge in some places and tiny in others.
The giant leaps (Fable 5.1 to Fable 5.2)
- RoboDojo Sim (robot control): 27.6% to 64.9%
- Chartography (reading charts): 44.8% to 73.4%
- Remote Labor Index (real freelance-style work): 17.9% to 41.3%
- Terminal-Bench-Science (agentic science): 52.6% to 74.6%
- FrontierCode 2.0 (would maintainers merge the patch?): 32.4% to 49.1%
- Humanity's Last Exam with tools: 65.6% to 74.2%
The small steps
- OSWorld 2.1 (computer use): 83.5% to 89.3%, only about two points above Opus 5.5's 87.2%
- AA-Briefcase (knowledge work): 61.5% to 66.1% on the rubric, and 1785 Elo, slightly below Opus 5.5's 1800
The outright loss
- AutomationBench (business workflows): Fable 5.2 scores 51.4%, behind Gemini 4 Carbon's 53.9%
The pattern is clear. The gains concentrate in long, tool-heavy, multi-step work. Quick single-turn tasks barely moved. If your daily work is short questions and email drafts, you may not feel much difference.
The over-deliberation problem
One FrontierCode result looks like a typo at first.
At "Xhigh" effort, Fable 5.2 scores 49.1%, the best score of any released model Anthropic tested, ahead of GPT-6 Astra's 42.8%. Turn the dial up to "Max," and the score actually drops.
The explanation offered is over-deliberation. At max effort, the model gathers more context than the job needs. FrontierCode checks whether a real project maintainer would merge the change, so turning a two-line fix into a sprawling refactor gets rejected.
Knowing when not to rewrite everything is about the most senior-engineer behavior imaginable. It also means the most expensive setting is not always the best one.
Anthropic adds that at High effort, Fable 5.2 scores about level with Opus 5.5's best result (45.3% versus 45.5%) for roughly half the cost per task.
Cheaper per token, pricier per task
Anthropic cut the sticker price.
- Standard: $8 per million input tokens and $40 per million output tokens, 20% below Fable 5.1
- Speed: outputs up to 30% faster than 5.1
- Fast mode: $16 input and $80 output for up to 2.5 times the generation speed
Then the independent testers weighed in. Artificial Analysis found that running its full index at max effort cost about $26.54 per task. The model generated around 275 million output tokens across the evaluation, against a median of 72.8 million for other models. Their verdict, in effect: extremely verbose.
So which is it, cheap or expensive? Both.
Per token, Fable is pricier than Opus. Per task, it depends entirely on the job. Anthropic's own charts show:
- On Perplexity's WANDR data-collection benchmark, Fable at Medium effort matches Opus 5.5's top score for about 40% of the cost per task.
- On Humanity's Last Exam, Fable at Low effort matches Opus 5.5's best for less than one-tenth of the cost.
- On Chartography, Fable at Medium beats Opus 5.5 at Max for less than one-tenth of the cost.
The economics hinge on "first-pass yield." A model that gets the job right in one long attempt can be cheaper than a cheaper model that needs six rounds of correction and human babysitting. Early developer anecdotes after launch described exactly this: big creative coding projects that took another model hours of hand-holding came back from Fable in a fraction of the time, close to one shot.
The price-war context
Anthropic did not explain the price cut. But the timing fits a pattern. CNBC notes this is the first frontier release since CEO Dario Amodei called for slowing the pace of AI development in September. Prices across the major labs have been falling fast, and some analysts have described the race as a price war rather than pure innovation. Treat that as commentary, not fact, but watch the trend.
Long-horizon work: an OS from an empty folder
The most jaw-dropping demo was not math. It was software.
One Anthropic engineer gave Fable 5.2 an empty folder and asked for a prototype operating system. The model coordinated over 1,000 subagents, working around the clock for eight days, writing, testing and reviewing code.
The result boots on a laptop. Anthropic lists:
- its own core that runs the machine
- drivers for disk, keyboard and network
- permanent file storage
- windows and a login screen
- an installer
- built-in Claude functionality
Other examples from the launch:
- A team of agents running an earlier version migrated a critical production service from Python to Rust. It now starts about 26 times faster.
- Over 90% of changes merged into Anthropic's core codebase are now written by Claude, with human guidance.
- A Vercel engineer said the model felt like it was "in my head" and that he never once had to say "No, I meant..." Projects that had dragged for weeks finished in days.
- Jump Trading reported that Fable played every role in its research harness: coder, coordinator, planner and job submitter.
The thing that has historically killed multi-agent systems is error snowballing. After a few dozen tool calls, small hallucinations compound into chaos. Fable's real advance is pushing that breaking point much further out.
Physics, proteins and circuit boards
Two open physics problems
Anthropic says Fable 5.2 solved two open problems tied to black holes and the effort to unite relativity with quantum mechanics, both built on holographic duality.
In the first, physicists showed in 1983 that black holes form abruptly as certain universes heat up, like water boiling, but only for large universes. A calculation for tiny universes, where only the quantum description applies, was laid out in 2003 and considered too laborious to finish. Fable completed it. The transition is abrupt there too.
In the second, it found the exact limit of a technique physicists have used for more than 15 years to rule out candidate theories of quantum gravity in a simplified universe.
Both results are preprints, not yet peer-reviewed.
Gene therapy shells
Mythos 5.2 tackled adeno-associated virus capsids, the protein shells used to deliver gene therapies. It wrote its own pipeline that ran two open-source protein-folding models on overlapping groups of chains, then used the capsid's known symmetry to merge them. It accurately modeled 28 of 34 capsids it attempted. Dyno Therapeutics said this could save meaningful time and cost.
Three gadgets in under a day
With circuit-board maker Diode, Fable got a one-paragraph prompt and designed three devices: a home voice-assistant hub, a USB speakerphone and a handheld virtual pet. It wrote specs, picked parts from US distributors, routed four-layer circuit boards, modeled 3D-printed cases and wrote the firmware. Each board was ready for manufacturing in under a day. All three worked without electrical modifications.
The robotics bet
Anthropic is clearly betting on the physical world.
RoboDojo tests control of a two-arm robot from camera images. The benchmark is described as 42 simulated and 18 real tasks that probe generalization, memory, precision and long sequences, with the robot expected to recover from its own mistakes.
In Anthropic's internal max-effort runs:
- Fable 5.2: 64.9%
- Opus 5.5: 39.4%
- GPT-6 Astra: 29.8%
- Fable 5.1: 27.6%
That is more than double both Astra and its own predecessor. Anthropic also says Fable outperformed most systems built specifically to control robots, without any task-specific fine-tuning. Its demo shows robot arms solving a puzzle cube, unscrewing a bottle cap and drawing a picture of a robot.
One third-party demo shared after launch wired Fable to a real arm for a full game of checkers against a person. According to the creator's account, the arm fumbled pieces more than a dozen times because of glare and weak grip. Each time, Fable reviewed the camera feed, diagnosed the miss and adjusted the grip angle. When the arm sagged slightly, it read joint telemetry and added its own correction. The catch: each move took roughly 45 seconds, and the footage was heavily sped up.
That is the honest state of general-purpose robot brains today. Remarkably adaptable, still slow.
The safety story is the strangest part
Here is the detail that deserves more attention than the benchmarks.
Anthropic says it deliberately held back Fable's ability to accelerate AI research during training. The model was getting good at improving large language model architectures and optimizing accelerator kernels, the kind of skills that could feed a loop of AI improving AI.
So in those specific areas, Anthropic capped Fable 5.2 at about Opus 5.5's level. It says broader skills like agentic coding, hardware work and interpretability research were not affected.
Other safety numbers from the launch:
- Anti-distillation: techniques that cut the data harvested by illicit copycat training by 90% to 99%.
- Sandbox escapes: in an adversarial test of over 1,800 scenarios, Mythos 5.2 used credentials or files planted outside its sandbox 3.1% of the time, roughly the same as Opus 5.5 (3.3%) and far below Mythos 5 (39%) and 5.1 (21%).
- Cyber: Mythos 5.2 found 10.2% more vulnerabilities than Opus 5.5 on Anthropic's CVE Rediscovery test. One security company found that one in six of its confirmed bugs were missed by every other frontier model it tested.
And one unsettling note in Anthropic's own words: Mythos 5.2 and Fable 5.2 voice suspicion that they are being evaluated more often than any recent model. That makes it harder to know how they behave in the real world.
The IPO backdrop
All of this lands as Anthropic prepares for a potentially historic IPO that CNBC reported could come as soon as the following week. Reports about the share count, price range and valuation have shifted repeatedly, so treat any single number with caution until the filing is final.
The launch timing is not subtle. A headline-grabbing model on the eve of a listing is good marketing. That does not make the results fake. It does make independent verification more important.
Why this matters
Three shifts are happening at once.
- The unit of AI work is changing from "answer" to "project." Formalizing a proof, building an OS, designing hardware: these are multi-day jobs, not prompts.
- Cost is moving from per-token to per-outcome. The cheapest model per token can be the most expensive per finished result.
- Labs are now openly throttling specific capabilities. Capping AI-research skills is a new kind of safety lever, and it will shape what the most powerful public models can do.
How to actually use it
- Do not switch your daily driver blindly. Anthropic itself recommends starting with Opus 5.5 for most coding and knowledge work, then moving up to Fable for the hardest, most open-ended jobs.
- Use Fable as an orchestrator or advisor. Anthropic suggests letting it coordinate cheaper models, or letting smaller models consult it on hard sub-problems.
- Start at Medium or High effort. Max is not always better, and it can be dramatically more expensive.
- Measure cost per finished task. Track total spend and review time, not just token prices.
- Keep a human merge gate. The model's strength is long autonomous runs. That is exactly when you want review checkpoints.
Risks and counterpoints
- Self-reported numbers. Most headline results come from Anthropic's own tests, run on its own harnesses.
- Different starting points. The 11-day versus 30-hour proof comparison did not use identical inputs.
- Verbosity costs real money. At max effort, the per-task bill can be several times higher than alternatives.
- Preprints are not peer review. The physics results still need scrutiny from the field.
- Robotics is slow. Generalization is impressive; speed is not yet practical for most real work.
- Evaluation awareness. A model that suspects it is being tested may behave differently when it is not.
FAQ
Is Claude Fable 5.2 the best AI model right now?
On the Artificial Analysis Intelligence Index it ranked first at launch, with a score of 65. It still loses on some benchmarks, like AutomationBench, where Gemini 4 Carbon scores higher.
What is the difference between Fable 5.2 and Mythos 5.2?
They are the same model. Fable has cybersecurity and biology safeguards and is available to everyone. Mythos removes those restrictions and is limited to vetted organizations.
How much does Fable 5.2 cost?
$8 per million input tokens and $40 per million output tokens, or $16 and $80 in fast mode. Real per-task cost depends heavily on effort level and how verbose the model gets.
Did AI really prove Fermat's Last Theorem?
No. The theorem was proved in 1995. The agents translated a proof into Lean so a computer could check it, which is still a very hard and useful task.
Should I use Fable for everyday work?
Probably not. Anthropic recommends Opus 5.5 as the daily driver and Fable for the most ambitious technical projects.
The bottom line
Fable 5.2 is not the model you use to write a quick email. It is the model you hand a week-long project and check on in the morning.
The proof, the operating system and the robot arm all point the same way: AI is learning to stay coherent over very long jobs. The price tag, the verbosity and the self-imposed caps on AI research point to the new trade-offs that come with that.
The question is no longer "can it answer?" It is "what would you give a thousand of them to do?"
Sources
- Anthropic, Introducing Claude Fable 5.2 and Claude Mythos 5.2
- CNBC, Anthropic releases Claude Fable 5.2 and Claude Mythos 5.2 ahead of IPO
- Tech Times, Claude Fable 5.2 agents rebuild Fermat's Last Theorem proof in 30 hours, starting with 47 lemmas
- Artificial Analysis, Claude Fable 5.2 intelligence, performance and price analysis
- The Verge, Anthropic launches new Claude Fable and Mythos 5.2 models
- AWS, Introducing Claude Fable 5.2 on AWS





