Anthropic's new Fable 5.2 model posts record scores on math benchmarks (standardized tests used to compare models), but a more useful measure for working researchers is what it does with real, unsolved problems under ordinary usage limits. In one such test, Fable 5.2 worked for hours on a repository of more than 40 open PhD-level problems. It produced 542,941 lines of code, notes and audit data, plus a clean three-page report. The documented result is careful, audited progress rather than a finished solution to the hardest open cases, and the weekly compute allowance ran out within about three to four hours.
What Anthropic reports about Fable 5.2
Fable 5.2 arrived shortly after Opus 5.5, Anthropic's previous flagship. The headline claim is a formalization of the proof of Fermat's Last Theorem, meaning the proof was rewritten in a form a computer can check step by step. According to Anthropic, 1,000 Fable 5.2 agents (independent copies of the model working in parallel) completed this in roughly 28 to 30 hours. The resulting proof is one-seventh the length of earlier computer-verified versions.
Anthropic ran Opus 5.5 on the same task. After 18 hours, it had reached 26% completion.
The scores on FrontierMath, an expert-level mathematics benchmark from Epoch AI, are as follows.
| Model | FrontierMath | FrontierMath Tier 4 |
|---|---|---|
| Fable 5.2 | 95.1% | 100% |
| GPT-6 Astra | 93.7% | 97.6% |
| Opus 5.5 | about 91% | not given |
Anthropic also reports that Fable 5.2 contributed novel solutions to two previously unresolved problems in theoretical physics.
A 1,000-agent run is not what a subscriber gets
The Fermat result used 1,000 agents running continuously for more than a day. Consumer Claude subscriptions cap parallel agents at 20 by default, and the top consumer tier costs $200 per month. Reproducing a run of that size on a personal account is out of reach, since the weekly allowance would be gone within hours.
That gap is why a test on a real research project is informative. It shows what the model does with the resources an individual researcher actually has.
The test: four years of open problems on random unitary matrices
The test material was a research directory built up over four years of doctoral study. It holds more than 40 open problems on the spectral properties of random unitary matrices. The specific questions concern asymptotic moments of the derivatives of characteristic polynomials in three regimes:
- inside the unit circle
- on the boundary
- under microscopic scaling near the boundary
The case of dimension N = 2 was already resolved in a doctoral dissertation. Dimension N = 3 and higher were open going in.

▲ Agents drafting, computing and auditing in parallel
How the run was set up
- An earlier model, Fable 5.1 Max, summarized the repository structure and drafted the starting prompt with its constraints.
- Fable 5.2 Max ran in Claude Code, Anthropic's coding and agent tool, pointed at the research directory.
- The prompt used the keyword ultracode, which triggers multi-agent parallel execution across sub-problems.
- The agents read the notes, drafted proofs, wrote Python scripts to check claims numerically and logged conclusions in Markdown, a plain-text note format.
What happened in the first hours
Activity was heavy from the start. Within 35 minutes, the agents had made more than 9,000 edits across repository files. Fable 5.2 updated the master README file, created sub-trackers for lemma verification and ran Python scripts to check algebraic identities numerically.
Compute went quickly as well:
| Elapsed time | Weekly Fable allowance used |
|---|---|
| About 1 hour | 53% |
| About 3 to 4 hours | 92%, then 100% |
Anthropic currently offers two free Fable weekly limit resets, which expire on March 9, 2027. With those, the work continued through the equivalent of three weeks of Fable quota.
What came out
Git, the version-control system that records changes to a project, logged 30 commits affecting 852 files. The added lines break down this way:
| File type | Lines added |
|---|---|
| JSON verification audit logs | 445,139 |
| Python | 70,565 |
| Markdown notes | 24,908 |
| Other files | 2,329 |
| Total | 542,941 |
At the end, Fable 5.2 compiled a status report in LaTeX, the typesetting system used for mathematical papers, as a three-page PDF. The report covers rigorous findings on characteristic polynomial moments for unitary matrices. It documents progress on a goal labeled KF-1, a planar representation at fixed N, with verification achieved for N = 2. It also records five rounds of independent audits across 16 reports. Those audits identified and corrected non-uniformity statements in the representation theorem.

▲ Large raw output against a three-page report
Where it advanced and where it stopped
Set against the starting point, the picture appears mixed in an instructive way. The verified milestone in the report sits at N = 2, the dimension that already had a resolution. The documented milestones do not include a completed proof for N = 3, the open target. On this reading, Fable 5.2 built a checked foundation and an audit trail around the known case, and corrected imprecise statements along the way. Nothing in the documented results shows the open problem closed.
Two limits stand out.
- Compute. One hour of multi-agent work consumed more than half a weekly allowance, and three to four hours consumed all of it.
- Review. Hundreds of thousands of lines are far more than one person can check line by line. The researcher ends up relying on the automated audit framework the agents themselves produce, which makes the design of that framework a central part of the work.
What changed from Opus 5.5
Opus 5.5 was already useful enough to be heavily integrated into a mathematical research workflow. The differences reported for Fable 5.2 fall into three areas.
- Scale of sustained work. On the Fermat formalization, Opus 5.5 reached 26% after 18 hours, while 1,000 Fable 5.2 agents finished in roughly 28 to 30 hours.
- Benchmark score. Fable 5.2 scores 95.1% on FrontierMath against about 91% for Opus 5.5.
- Writing style. Earlier Anthropic models sometimes drifted into speculative, quasi-metaphysical language when handling deep abstractions. Fable 5.2 keeps to direct, rigorous mathematics that can be audited.
Taken together, Fable 5.2 appears to be a major step forward in autonomous mathematical reasoning and in synthesizing scientific literature. That is a judgment from a single research project, not a general measurement.
Key points and what to do
Fable 5.2 can work through a real research repository for hours, write its own verification code, audit its own claims over several rounds and summarize the state of the work in a short, readable report. The documented results of this test do not include a solution to the hardest open case, and compute limits and the sheer volume of output remain real constraints.
Researchers who want to try a similar workflow can take these steps:
- Organize the project into separate Markdown tracking files with explicit audit and ground-truth criteria, so the agents can verify their own work.
- Use a strong model first to survey the repository and write a thorough starting prompt before launching an autonomous session.
- Have the agents write Python scripts that independently check identities and error bounds.
- Reserve the top model's allowance for core mathematical reasoning, and send routine coding and other side tasks to smaller, cheaper models.
- Check account usage settings for limit resets before starting a large run, and expect a multi-agent session to use a weekly allowance within hours.
- Read the final report and the audit summaries closely, since reviewing the raw output line by line is not realistic.


