Claude Fable 5.2 agents ran unsupervised for six weeks and solved two open black hole physics problems, including a calculation stuck since 2003. Here's how the setup actually worked.
Vamsi Tallapudi
Manager, Architect Technology at Cognizant
Claude Fable 5.2 spent six weeks running unsupervised in the background and came back with solutions to two theoretical physics problems that had stumped human researchers for up to 23 years. Anthropic published the results on November 17, 2026, and the part everyone is arguing about isn't the physics. It's that nobody told the agents to do this.
Here's the thing that makes this story different from the usual "AI does math" announcement: it started as a side project that nobody budgeted for.
In August 2026, Anthropic researcher Hongbin Chen was poking at Claude's capabilities, not running an official research program. He asked it to draw up a list of interesting open problems in theoretical physics, specifically around black holes and a technique called the bootstrap method. Hours later, Claude handed back 22 candidate problems.
Chen didn't pick the easiest ones and babysit them. He spun up a separate agent for each problem and let them run. The agents could read each other's progress posts and borrow ideas from agents working on related problems. Then he mostly left them alone for six weeks.
Two of those 22 threads produced real results:
Most AI-does-science stories involve a dedicated team, a curated benchmark, and months of scaffolding built specifically for the win. This one didn't. Chen's setup was closer to "ask a good question, assign the homework, check back later."
That's a different claim than "Claude is smart." It's closer to "autonomous agents left alone for weeks can self-organize around a hard problem without someone hand-holding every step." If that generalizes past physics, it matters a lot more than one Hawking-Page calculation.
I'd also point out the agents weren't working in isolation from each other, which is easy to skim past. Multiple agents independently redid almost every piece of the Hawking-Page calculation using different methods, and those independent derivations caught each other's mistakes. On the bootstrap problem, one agent noticed the pattern and a different agent built the actual proof technique around it. That's agent-to-agent division of labor that nobody explicitly orchestrated.
Three layers, and this is where Anthropic actually did its homework instead of just taking the win:
On top of that, the modular bootstrap result was reportedly reproduced independently using the publicly available Claude Opus 5.5, not just Fable 5.2. That's a meaningfully stronger signal than "trust our internal eval."
Still, both results are Anthropic-deposited preprints. Neither has cleared peer review yet, and Hacker News piled onto that fact within hours of the post going up.
Quite a bit, and I'd rather lay it out than pretend this is a clean win.
The two solved problems live in simplified, lower-dimensional toy universes, not the messy four-dimensional one we actually occupy. A physicist on the Hacker News thread for this story made the obvious point: there are countless open problems in theoretical physics sitting around, and plenty of them are tractable but just not that interesting to the field. Significance comes from peer review and community adoption over years, not from a company blog post. Nobody's handed Claude a Nobel-tier unsolved problem and watched it fall yet, and these two don't really qualify as that either.
Reproducibility is the second complaint, and it's a fair one. Anthropic hasn't published the actual prompts and agent transcripts that produced the results, which matters because LLM output isn't deterministic. Without that, outside researchers can't rerun the exact process and confirm it wasn't a lucky roll that got cherry-picked out of 22 parallel attempts.
And yes, the timing is suspicious. This landed as Anthropic pushes toward an IPO targeting a valuation in the $1 trillion to $1.5 trillion range, on the back of roughly $33.2 billion in nine-month 2026 revenue. A "our AI does real science" story is good for a roadshow slide regardless of whether it's also good science. Both things can be true at once: the math can check out in Lean and the announcement can still be timed for maximum investor attention.
Yes. This isn't some special research build. Fable 5.2 is the same model Anthropic shipped on November 9, 2026, at $8 per million input tokens and $40 per million output, a 20% cut from Fable 5.1. That same launch week brought a Fermat's Last Theorem formalization demo and an 8-day prototype OS build. For how that pricing stacks up against the rest of Anthropic's lineup, see our Opus 5 vs Fable 5 price and benchmark breakdown. The black hole work is a separate result from the same model family, published about a week after the launch, running on the same API anyone with a paid plan can hit today.
That matters because it changes the claim from "Anthropic's secret internal model is smart" to "a model you can currently rent by the token did this when given six weeks and no supervision." Whether that's reassuring or unsettling probably depends on how you feel about AI labs generally.
Honestly, the physics itself won't change your day. What should get your attention is the operational pattern: open-ended problem generation, parallel agent assignment, agent-to-agent collaboration, weeks of unattended runtime, then formal + human verification at the end. That's a workflow, and it's one you can borrow for anything that's hard to verify step-by-step but easy to check once it's done, think large refactors, exhaustive test generation, or combinatorial search problems in your own codebase.
The honest caveat: Anthropic had a physicist double-check the output and a formal proof checker backstopping the math. If your domain doesn't have an equivalent "Lean" to catch a confidently wrong agent, six weeks of autonomous drift could just as easily produce six weeks of autonomous nonsense that looks clean until someone who knows the field actually reads it.
It completed a calculation about the Hawking-Page phase transition that was set up by physicists in 2003 and left unfinished because it was too laborious to do by hand. Multiple Claude agents worked it independently, cross-checked each other, and the result was later formally verified in Lean and reviewed by a physicist.
It's the question of exactly when a hot gas cloud collapses into a black hole instead of staying as gas. Fable 5.2's result shows the transition is abrupt even in a fully quantum setting where classical general relativity doesn't apply, matching what physicists expected but couldn't previously prove.
Modular bootstrapping is a 15-year-old technique for ruling out candidate theories of quantum gravity in simplified universes. Claude proved an exact ceiling on the lightest possible black hole mass, M₁ ≤ c/Λ, with Λ approximately 9.1266, a provably optimal bound rather than an estimate.
About six weeks, starting in August 2026. Researcher Hongbin Chen assigned 22 candidate problems to separate Claude agents and largely left them alone, checking back periodically rather than actively directing the work.
Three ways: agents independently re-derived the same results by different methods, the final proofs were formally checked in Lean (a proof assistant) in September 2026, and a physicist reviewed every main result before Anthropic published it. The bootstrap result was also reportedly reproduced independently using Claude Opus 5.5.
No. Both results are posted as preprints and haven't cleared formal peer review. Hacker News commenters flagged this quickly, along with the fact that Anthropic hasn't published the raw prompts and transcripts needed to independently reproduce the process.
Yes, it's the general-release model that launched November 9, 2026 at $8/$40 per million tokens, not a special internal research build. See how Fable 5 compares to Opus 5 on price and benchmarks for more on where it sits in Anthropic's lineup.
New AI tools, automation workflows, and course drops — straight to your inbox. Join 2,400+ builders.
Gemini 4 Neon beats Claude Opus 5.5 on composite benchmarks (61 vs 59) at 25% lower price, but Opus 5.5 still wins long agentic coding sessions 86.9% to 79.8%.
CVE-2026-123007 is a CVSS 9.2 nginx HTTP/2 flaw going back to 2015. Here's how to check if you're exposed and patch to 1.31.9 or 1.30.7 today.
Claude Code, Codex CLI, Pi, and OpenCode each win on different models. Runta's eval found harness choice swinging cost up to 4.4x on the same model.
Comments