Claude Fable 5.2 solves two open black hole problems: the 2003 Hawking-Page problem and an exact modular bootstrap ceiling
🔎 Black holes, two open problems — and finally an artifact to believe in
On November 17, 2026, Anthropic announced that Claude Fable 5.2 had solved two open problems in theoretical physics: the Hawking-Page calculation, unblocked since 2003, and an exact ceiling for the simplest modular bootstrap. Not benchmark questions: problems the community had been stuck on for over twenty years.
What sets this announcement apart from the long series of "the AI supposedly did it" claims: everything is verifiable. The results passed a formal proof in Lean as early as September, the Hawking-Page calculation was recovered via two independent routes, and a physicist reviewed each main result. The preprints are public.
Context matters as much as the result. A few months after Claude Fable 5 was made available to the public, and after a series of summertime mathematical feats, Anthropic is scaling up: 22 candidate problems, agents launched in the background for six weeks. What follows breaks down what was proven, how — and what Anthropic does not claim, because that is where credibility is at stake.
The Essentials
- Hawking-Page, finally completed. The calculation identified as a bottleneck in 2003 by Ofer Aharony and his colleagues—a number requiring thousands of terms—has been brought to a conclusion: the abrupt gas $\to$ low-coupling black hole transition in N=4 super Yang-Mills is established.
- An exact ceiling, not an estimate. The simplest modular bootstrap is capped at $\Lambda \approx 9.1266$, or $c/9.1266$ — proven optimal for this version of the method, which cannot reach $c/12$, where black holes appear.
- A reproducible method. 22 candidate problems listed in August, one agent per problem, six weeks of execution without special setup, progress notes shared between agents.
- Verification did the heavy lifting. Independent duplication, comparison to published bounds, formal proof in Lean in September, human review.
- The limits are acknowledged. No result describes our universe, no practical application, and quantum gravity is not resolved.
Recommended Tools
To read, verify, or reproduce this type of result, the minimal toolkit consists of four lines.
| Tool | Primary Use | Price (November 2026) | Ideal for |
|---|---|---|---|
| Lean 4 + mathlib | Formal verification of proofs | Free (open source) | Certifying that an AI result is mathematically sound |
| Claude (Fable 5.2) | Long-term research agent sessions | From $20/month (November 2026, check anthropic.com) | Reproducing a "one agent per problem" setup |
| Preprints hexagonmath.org | Reading the two original papers | Free | Verifying statements and methods yourself |
| arXiv | Bootstrap literature (Hellerman, Hartman-Mazáč-Rastelli) | Free | Locating published bounds before accepting a "new" bound |
The Hawking-Page calculation: a 23-year roadblock finally cleared
Yes, the calculation identified as the blocker in 2003 has been carried through to completion. It's the first of the two results announced by Anthropic, and the most symbolic.
The story begins in 1983, when Hawking and Page describe an abrupt transition between a thermal gas and a black hole in anti-de Sitter space — the canonical image is that of the geometry itself boiling (The AI Gazette, November 19, 2026). An extension of this result follows in 1999.
Then, in 2003, Ofer Aharony and his colleagues identify the bottleneck: a number that requires thousands of terms to evaluate. Too tedious to do by hand, never finished — the calculation sat there for more than twenty years.
Claude Fable 5.2 has completed it. Concretely, the abrupt gas → black hole transition at weak coupling in N=4 super Yang-Mills is now established (CellCog, November 18, 2026). A problem open for 23 years falls.
Two methods, one answer
Validation is the real story. "Weeks after one of the agents reached a solution, a physicist, who had independently used Claude to tackle the same problem, shared his result," reports Anthropic. Two solutions, obtained by different methods, that validate each other.
This is exactly the protocol you'd want to see with every announcement of this kind: two independent paths converging, not a model correcting itself. And the detail that this physicist was already using Claude of his own accord, without any coordination with Anthropic, says a great deal about actual adoption in the field.
Modular bootstrap: from ~c/9.1 "suggested" to c/9.1266 proven
The second result is more technical, and more rigorous: Anthropic doesn't just improve a numerical bound, it proves the exact ceiling of the method.
Let's recap the trail. The ceiling of the simplest modular bootstrap was fixed at c/6 by Simeon Hellerman, then lowered to c/8.5 by Hartman, Mazáč and Rastelli. Numerical studies suggested around c/9.1 — without proof (The AI Gazette).
Claude establishes Λ ≈ 9.1266, i.e., an exact ceiling of c/9.1266, using a spectral curve — a tool never before applied to this problem. And the bound is proven optimal for this version of the method: it cannot reach c/12, where black holes appear (codingislove, November 18, 2026).
In other words, we don't just know "where the method stands": we know precisely where it stops. Moving from "the numerics suggest" to "it's proven, and it's the ceiling" is the difference between an indication and a result.
The path there was bumpy. A Claude agent had already attempted the problem — and failed after 61 parallel attempts. The breakthrough came from elsewhere: a pattern spotted in the literature by one agent, turned into a method by another who was working on sphere packing. Both preprints are public on hexagonmath.org.
The methodology: 22 problems, six weeks, zero special setup
No, no exotic infrastructure was required. In August, Anthropic researcher Hongbin Chen simply asked Claude to list open problems in physics: 22 candidates came out of the exercise, and each one got its own agent.
These agents ran "with no special setup for about six weeks," notes codingislove. No in-house fine-tuning, no custom tooling: just long-running execution in the background.
They were granted a single freedom, and it probably decided everything: reading the other agents' progress notes. The bootstrap result is the direct product of that — a pattern found by one agent, shaped into form by another.
In the end, Claude wrote the paper "with minimal guidance," and a physicist verified every main result (Anthropic, November 17, 2026).
The architecture is almost disappointingly simple: persistent agents, shared documentation, time. No sophisticated orchestrator. Good news for anyone who believes agentic research requires an exotic stack — what it requires above all is patience and traceability.
The timeline, in four dates
It all played out over four months, between the drawing up of the list in August and the November 17 announcement.
| Date | Milestone |
|---|---|
| August 2026 | Hongbin Chen has Claude list 22 open problems; one agent is assigned to each |
| August – September 2026 | ~6 weeks of background execution, shared notes; 61 parallel attempts fail on the bootstrap before it clicks |
| September 2026 | Formal verification of the results in Lean |
| November 17, 2026 | Anthropic's announcement; public preprints on hexagonmath.org |
Independent analyses followed within 72 hours: codingislove and CellCog on November 18, The AI Gazette on the 19th, The Trillioniar on the 20th. A story read as methodological as it is physical — that's a good sign.
Lean as the standard of proof: the real lesson of the announcement
The most durable result is not in the physics papers, it is in the validation protocol. "Verification did the heavy lifting," summarizes codingislove.
Four layers, in order: independent duplication (Hawking–Page recovered via two routes), comparison against published results (c/8.5, ~c/9.1), formal proof in Lean in September, human review of each main result.
None of these layers requires trusting Anthropic. That is the essential point: a company that sells tokens obviously has an incentive to tell this kind of story, and skepticism must remain the default regime. A Lean artifact, on the other hand, can be verified without believing anyone.
After years of contaminated benchmarks and cherry-picked demos, this is the only standard that holds up. Any announcement of the form "AI proved X" should come with its formal proof or its independent duplication — otherwise, it's marketing.
Moreover, this is not an isolated fire. This summer, Claude had extracted 663 unknown theorems from Gödel's notebooks, and Google Teamwork had proved the Courtade–Kumar conjecture with agent swarms. The difference here: formal verification is at the heart of the announcement, not in an appendix.
What this changes for physics — and what doesn't change
Let's start with what doesn't change: none of this describes our universe. Anthropic states this explicitly — no result has any practical application, and no one claims to have solved quantum gravity (CellCog). A rare clarification in an AI announcement, and one that deserves credit.
The playground is N=4 super Yang-Mills, a remarkably symmetric toy model — not our cosmos. The established Hawking-Page concerns a weak-coupling regime: mathematically clean, physically unrealistic.
What changes is the internal economy of the discipline. The tedious calculations — those numbers running to thousands of terms —, the numerical scans of bounds, the consistency checks: all of it becomes delegable to agents.
The profession moves up the chain: less manual calculation, more problem selection, direction, and verification. The independent physicist who was already attacking Hawking-Page with Claude, with no connection to Anthropic, shows that adoption is already underway.
My prediction: the bottleneck shifts from computational capacity to choosing the right problems. Knowing which problem deserves six weeks of agents becomes the rare skill — not knowing how to set up the calculation.
And for devs: three takeaways you can apply tomorrow
You don't have black holes to solve, but the protocol itself carries over as-is to software research.
- Decompose. 22 independent problems rather than one mega-agent: natural parallelism, isolated failures, easy restarts.
- Share notes. The bootstrap discovery was born from cross-pollination between agents; your agents should read each other's work too.
- Verify before believing. Duplication, comparison against known results, formal proof, human review — for code, this translates into specifications, formal methods, and property-based testing.
On the choice of engine, let's keep a cool head. Agentic leaderboards put GPT-5.5 on top (98.2), ahead of Gemini 3 Pro Deep Think (95.4) and Claude Opus 4.7 Adaptive (94.3) as of mid-2025 — but no benchmark score replaces a verifiable artifact, and that's the whole Lean lesson. To choose based on your use case, our comparison of the best LLMs for coding remains the best entry point.
On the other hand, six weeks of continuous execution — that's something to sandbox. The zero-click RCE Plugin4Shell went through Claude Code, Codex, Copilot, and Gemini CLI — a long-running agent with network access and literature reading is as much an attack surface as a collaborator. Isolate, log, review.
Last point, less technical: don't build your entire pipeline on a single vendor. The episode of the blocking of Claude Fable 5 and Mythos 5 — models disabled for all users after a Trump order — showed that an agent pipeline can be cut off overnight, for reasons unrelated to the technology.
❌ Common Mistakes
Mistake 1: Believing quantum gravity is solved
No, and Anthropic doesn't claim it. N=4 super Yang-Mills is a toy model; no result describes our universe, and no practical application is being announced. The caution is part of the announcement — read that too.
Mistake 2: Confusing a method's ceiling with a physical frontier
c/9.1266 is optimal for this version of the bootstrap, not a law of nature. The method cannot reach c/12, where black holes appear — another tool could push further. The ceiling bounds the tool, not the world.
Mistake 3: Sharing the announcement without the artifact
The preprints are public (hexagonmath.org) and the Lean proof exists. Any "the AI proved X" announcement that provides neither a formal proof nor independent replication should be treated as marketing, not science.
Mistake 4: Copying the "six weeks without supervision" without guardrails
What works on pure mathematics — where no side effects are possible — is dangerous on real systems. Sandbox, minimal permissions, logs, human review: otherwise, re-read the Plugin4Shell episode before launching your agents.
❓ Frequently Asked Questions
Did Claude Fable 5.2 solve these problems on its own?
No. A researcher, Hongbin Chen, framed the list of 22 problems, agents ran for about six weeks, a physicist reviewed every main result, and the Lean proof certified the whole thing. The AI did the computing; the human did the science. Without that framing, none of these discoveries would have happened.
Does the Hawking-Page calculation concern real black holes?
Not the ones in our universe. It takes place within the N=4 super Yang-Mills framework, a toy model. The abrupt gas → black hole transition is now established at weak coupling — a mathematical consistency result, with no practical application announced by Anthropic.
What is the modular bootstrap, in simple terms?
A method that reconstructs a physical theory from its consistency constraints alone. The "ceiling" measures how far its simplest version can push the analysis: Anthropic proves that it stops exactly at c/9.1266, before c/12 where black holes appear.
Why is the Lean proof so important?
Because it turns a claim into a verifiable artifact: a proof assistant checks every step, trusting no one. Since September, both results have been formally certified. It's the most serious antidote to the unverifiable AI announcements that have been piling up for the past two years.
Can the experiment be reproduced?
Both preprints are public on hexagonmath.org and the methodology is detailed by Anthropic and codingislove. But plan on six weeks of continuous execution and a suitable subscription. Start with smaller problems: it's the method that matters, not the scale.
✅ Conclusion
For the first time, an "AI has solved an open problem" announcement comes with what it takes to be believed: a formal proof, an independent replication, and human review — and that's the real news of November 17. Fancy launching your own agentic sessions? Our Claude vs ChatGPT comparison helps you pick the right engine for your next six weeks.