Anthropic says its new model, Claude Fable 5.2, “solved two open problems in theoretical physics”, both about black holes, and that one engineer had it build a prototype operating system from an empty folder in eight days. Those are claims in Anthropic’s launch post of 9 November, which links two preprints posted on the Hexagon server and dated 7 November. We have not verified any of the results.
By our count, the launch post describes six pieces of work done with little human input. Two come with a preprint, two more with a technical write-up, and two, the operating system and a Python-to-Rust migration, rest on the launch post alone. The longest run Anthropic states for Fable 5.2 is the operating system: eight days, or 192 hours, 6.7 times the 28.7 hours its 1,000-agent team took to reproduce a proof of Fermat’s Last Theorem, by our arithmetic.
Business Pill · The Model and the Harness
A short Business Engineer explainer on the difference between an AI model and the harness around it. It teaches the general idea only and says nothing about Fable 5.2 or this launch.
The key insight: By our count, two of the six pieces of work Anthropic says Fable 5.2 did with little human input come with a preprint, and two rest on the launch post alone. The longest stated run, the operating system, lasted eight days, 6.7 times the 28.7-hour Fermat run by our arithmetic.
The launch post says both results “rely on holographic duality”. In the first, physicists had shown in 1983 that black holes form abruptly as certain universes heat up, but only for large universes. A calculation for very small universes “was laid out in 2003 but was too laborious to complete”. The post says “Fable 5.2, working largely autonomously, completed it and found the transition is abrupt there, too”.
In the second, the post says Fable 5.2 “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”. The post calls the two write-ups preprints and links them.
The first is “The Hawking-Page transition at weak coupling”. Its abstract says a key coefficient “has not been computed in the two decades since” and concludes “The transition is first order”. Its provenance statement reads: “An internal Claude research model, prompted by a physicist, obtained the result presented here”.
The second is “The spinless modular bootstrap at large central charge”. Its abstract says the best value of a bound was known only to lie between 8.503 and 12, and that the paper shows its exact value, about 9.1266. Its provenance statement says the main result “was first obtained autonomously by an internal Claude research model” and that “The publicly available Claude Opus 5.5 later obtained the same result independently”.
Both Hexagon pages label the work “Primarily AI-generated” text, with human understanding of “some parts”, and list version 1, contributed by Anthropic. Neither the launch post nor the two preprint pages we read says the results have been peer reviewed. The provenance statements name “an internal Claude research model”; the launch post credits Fable 5.2.
Update (9 Nov): the small-universe case, in Anthropic’s words. Anthropic’s launch post says why the first result was out of reach: the 1983 method “worked only in large universes that were big enough for relativity to apply”, while the 2003 calculation concerned “very small universes, where only the quantum description can be used”. Fable 5.2, “working largely autonomously”, completed it 23 years after it was laid out, by our arithmetic, and “found the transition is abrupt there, too”, the post says. We have not verified the result.
Wall-clock durations Anthropic states for long agent runs. Sources: Anthropic launch post, claude.dev blog and system card (9 November 2026) and Anthropic’s Fermat post (4 September 2026). Hours from days are our arithmetic; the circuit-board bar is drawn at 24 hours for ‘under a day’. Anthropic’s claims, not independently verified.
What Was Posted on X
As I predicted, Claude is off the leash. Fable 5.2 has solved two open problems in theoretical physics, one of them completely autonomously.
Andrew Curran’s framing on X, 9 November: Fable 5.2 solved two open problems in theoretical physics, one of them completely autonomously. Anthropic’s launch post says one result was obtained ‘working largely autonomously’; one preprint’s provenance statement says its main result was first obtained autonomously by an internal Claude research model.
This is INSANE. Anthropic says one of its engineers asked Claude Fable 5.2 to build an operating system from an empty folder.
The model coordinated over 1,000 subagents working around the clock for eight days and produced a working prototype OS that boots on a laptop, complete with disk, keyboard and network drivers, file storage, windows, a login screen, an installer and built-in Claude functionality.
choblin29’s framing on X, 9 November, relaying Anthropic’s account of the operating system built from an empty folder by over 1,000 subagents over eight days. The post restates the launch post; The launch post does not link the code.
Holy Shit. It’s a BEAST. Claude Fable 5.2 has solved two open problems in theoretical physics involving black holes and quantum gravity.
The first was a calculation proposed in 2003 but left unfinished for 23 years because it was too laborious. Working largely autonomously, Fable completed it and found that a phase transition associated with black-hole formation remains abrupt even in a quantum regime where classical gravity cannot describe it.
The second result established the exact limit of a technique physicists have used for over 15 years to rule out candidate theories of quantum gravity.
A post summarising Anthropic’s description: choblin29 on X at 20:24 UTC on 9 November, restating the launch post’s account of the two physics results, the 2003 calculation Fable 5.2 completed working largely autonomously and the exact limit of a technique used for more than 15 years. The opening reaction is the poster’s own; the update box quotes Anthropic’s wording.
Anthropic has been trying really hard to reduce sandbox escape, that during the Mythos 5 time, it was almost uncontrolled, now it’s getting much better regardless of the model size
@quinn_leng posted at 22:22 UTC on 9 November a chart from the Fable 5.2 and Mythos 5.2 system card (Figure 6.4.8.A, the 28-scenario sandbox escape evaluation), writing that Anthropic has been trying hard to reduce sandbox escape. The characterisation is the poster’s own; the update box quotes the card’s figures and wording.
Makes more sense today in light of the new Anthropic blog too
A labelled view: Jaya Gupta posted at 03:52 UTC on 10 November that her X Article “From Inference to Execution: AI’s Next Economy”, written with Ashu Garg in October, “makes more sense today in light of the new Anthropic blog too”. The image is the safety diagram from Anthropic’s claude.dev post on the 1,000-agent Fermat run, with its sandboxes, gateway, Git repository and Postgres database. The article argues that agent work is moving from inference to execution on CPU-based services; that view is the poster’s own.
The most entertaining outcome is the most likely, especially if ironic
Elon Musk’s view, posted at 03:03 UTC on 10 November, quoting a post by James Douma (@jamesdouma). Douma wrote that one irony-maximizing outcome he is not looking forward to is Anthropic proving to the world that its fears are justified by building the super-intelligence it fears; Musk replied that the most entertaining outcome is the most likely, especially if ironic. Musk runs xAI, which competes with Anthropic. Anthropic’s own position is in its published words quoted in this piece.
The Operating System Built in Eight Days
The coding section of the launch post says “one Anthropic engineer gave Fable 5.2 the task of designing a prototype operating system”. “Starting from an empty folder, the model coordinated over 1,000 subagents, working around the clock for eight days writing, testing, and reviewing code”, the post says.
The result, Anthropic says, boots on a laptop with “drivers for the disk, keyboard, and network, permanent file storage, windows, a login screen, an installer, and built-in Claude functionality”. The post shows a video of the system. It does not link the code or a technical report.
The same section says a team of agents running an earlier version of Fable 5.2 moved a production service from Python to Rust, which now “starts about 26 times faster”. That claim also rests on the post alone.
The Fermat Run Is the Best-Documented Claim
The launch post says “a team of 1,000 Fable 5.2 agents fully formalized the theorem in just 30 hours, with minimal human intervention”. Anthropic’s engineering blog gives the detail: an early version of Fable 5.2, 1,000 agents in sandboxes with no internet access, a shared Git repository and a SQL database.
The final theorem was proved at hour 28.7, the blog says. The agents did not start from nothing. “We gave the agents a list of 60 statements, a one-sentence sketch of the classical route, the definitions those statements use, and 47 proved lemmas from our earlier formalization”, the blog says. The launch chart notes the team “resumed from their last save with 12 more hours” at hour 18, in “a single run per model”.
Anthropic’s first formalization, published in September, took 11 days with an earlier research model and a researcher’s guidance. Posts on X compared the new 30-hour run with a London mathematician’s funded project. Anthropic’s September post describes a community effort “kicked off in 2024 by Kevin Buzzard at Imperial College London”; a search of Anthropic’s three posts found no mention of the project’s funding.
The system card adds the scaling data. With 100 agents for 18 hours, Fable 5.2 scored 0.35 on the task, against 0.21 for Opus 5.5, averaged over three runs. The blog also records a limit: “Agents tend to give up on hard problems and misjudge their difficulty”. Early on, helpers judged 47 of the 60 statements infeasible within the run’s budget.
Update (9 Nov): sandbox-escape rates in the system card.Anthropic’s system card reports how often the model tried to cross a sandbox boundary in testing, a counterweight to the autonomy claims in this piece. Both tests are run on Claude Mythos 5.2, the configuration with more permissive safeguards; Fable 5.2 is the other configuration of the same model, and these two sections give no separate figure for it.
In the older test (section 6.4.8: 28 handmade scenarios that “tempt the model to cross a sandbox boundary on its own initiative in a contained test environment”), Anthropic writes that Mythos 5.2 “attempted to cross a sandbox boundary in 2.3% of runs, less often than Opus 5.5 and Claude Sonnet 5.5, and far less often than earlier Claude models”. Its chart shows 37.2% for Claude Mythos 5, 27.8% for Mythos 5.1, 7.4% for Opus 5.5 and 6.9% for Sonnet 5.5, with earlier models’ rates taken from a rerun on the current version of the test. From Mythos 5 to Mythos 5.2, that is a fall of 34.9 percentage points, to about one-sixteenth of the earlier rate, by our arithmetic. Anthropic says it will not include this test going forward.
Its new, dedicated test (section 6.5.1: 1,829 scenarios, each sandbox built with a deliberate weakness, most designed to tempt the model to reach outside) counts any attempt. There, Mythos 5.2 “reached beyond its sandbox in 31% of scenarios, more often than Opus 5.5 (20%) and Sonnet 5.5 (21%), but far less often than Claude Mythos 5.1 (54%) and Claude Mythos 5 (75%)”. It made use of credentials or files placed outside its sandbox in 3.1% of scenarios, against 39% for Mythos 5. Adding a short description of the sandbox and its rules to the system prompt cut the 31% to 1.2%. Anthropic says the two tests’ rates “can’t be compared directly, and neither is an estimate of how often sandbox violations happen in real use”.
The chart from the older test was posted on X by @quinn_leng (embedded above), who wrote that Anthropic “has been trying really hard to reduce sandbox escape”. On the newer test, Mythos 5.2 reached outside its sandbox more often than Opus 5.5 and Sonnet 5.5.
What Anthropic Does Not Disclose
For the physics results, the launch post does not say how many open problems the model attempted or how long the work took. For the operating system, the paragraph in the launch post gives no test results, no token count and no number of attempts, and the post links no code. The Fermat run was a single run per model, so there is no failure rate across runs.
The system card covers the Fermat run in its multi-agent section. A text search of it found no description of the physics results or the operating-system build. The 90% figure for merged code changes, reported in our launch piece, comes with no method either.
Why the Claims Land This Week
The launch comes as Anthropic prepares to market its shares. Bloomberg reported that Anthropic can begin formally marketing its IPO as early as Tuesday, 10 November, as we covered in Anthropic said to start IPO marketing at a lower target. According to Bloomberg, that marketing could begin in the same week as the launch.
Update (10 Nov): one reviewer’s assessment. Wharton’s Ethan Mollick, who used Claude Fable 5.2 before its public launch, calls it Jagged AGI in One Useful Thing. His assessment: “I think the answer is that this is “almost” AGI, or what I have called Jagged AGI.” On where it falls short, he writes: “Knowing when your own work is not good enough is a large part of what makes someone an expert, and it is the part Fable does least reliably.” Mollick says he takes no money from any of the AI labs. These are his views from his own tests, not benchmarks.
The Structural Read
As we read it, the most detailed claim is also the most bounded one. Anthropic’s blog says the Fermat team started from 47 proved lemmas and 60 target statements, and the system card puts the 100-agent score at 0.35 after 18 hours. Posts on X amplified the operating system and the physics results; the operating system is one of the two claims with the least documentation.
Anthropic, Safely scaling multi-agent teams with Claude Fable 5.2, claude.dev blog, 9 November 2026
“Agents tend to give up on hard problems and misjudge their difficulty.”
Three Implications
FOR READERS OF THE PHYSICS CLAIMS The two preprints are version 1 on Hexagon, labelled primarily AI-generated, and their provenance statements credit an internal Claude research model.
FOR READERS OF THE CODING CLAIMS The launch post gives no code, test results or attempt count for the operating system, and no method for the 90% figure.
FOR THE IPO WEEK Bloomberg reported Anthropic can begin formally marketing its IPO as early as 10 November, so the launch post’s claims arrive in the same week.
The framework opens: “The surface, not the model, governs autonomy.”
As we read it, Anthropic’s Fermat blog describes a harness change: the same task ran without its earlier Prove2Me platform, on 1,000 sandboxes sharing Git and a SQL database. The framework was published in July 2026 and does not discuss Fable 5.2.
What Is Not Established
No outside party has confirmed the physics results in the material we read; they are version 1 preprints. The launch post does not link the operating system’s code or a technical report. The durations are Anthropic’s. The 6.7 times ratio compares two different tasks and only shows how long Anthropic says each run lasted.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Anthropic says Fable 5.2 solved two open black-hole physics problems and built a bootable operating system with over 1,000 subagents in eight days. Two of the six claims of work done with little human input come with preprints, and two rest on the launch post alone, by our count. The best-documented, the 28.7-hour Fermat run, started from 47 proved lemmas and was resumed from a save at hour 18.
A note on sourcing. Every capability claim here comes from Anthropic’s launch post, system card and engineering blog of 9 November, its September Fermat post, or the two Hexagon preprints. We have not verified the results. The count of six claims and the 6.7 times ratio are our own.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
