What Claude Did for Science in September 2026: Fermat, a CRISPR-like Enzyme, and More

The month at a glance

Mathematics: a machine-checked proof of Fermat's Last Theorem

Fermat's Last Theorem says no three positive whole numbers satisfy aⁿ + bⁿ = cⁿ for any n greater than 2. Andrew Wiles proved it in 1995, in a proof so long and specialized that few people can check it end to end. On September 4, 2026, Anthropic reported that Claude had turned that proof into a formal proof in Lean, a programming language whose checker verifies every logical step.

• Time: 11 days, working largely on its own.
• Size: about 13 million lines of Lean, roughly five times the size of Mathlib, the main shared library of formal mathematics.
• Work: about 29,500 intermediate theorems proved, using around 6 billion output tokens.
• Model: an internal research model roughly comparable to Claude Fable 5.1.
• Checks: the proof relies only on Lean's three standard axioms, with no unproven placeholders, and a second, independent proof checker confirmed all of its declarations.

Dozens of Claude agents worked in parallel on a shared dependency graph of theorems, so each agent knew what was already proven and what it could build on. Human input was limited to occasional high-level direction. Kevin Buzzard of Imperial College London, who leads the main human effort to formalize Fermat's Last Theorem, called it an extraordinary achievement.

The wider payoff is trust at scale. Modern mathematics rests on long chains of results that each depend on earlier papers, and a mistake deep in that chain can sit unnoticed for years. If formal proofs become cheap to produce, referees could ask for one alongside a paper, and libraries of verified results could grow far faster than human volunteers can manage. Buzzard described the result as a big step toward automatically formalizing the modern mathematical literature. The same idea, pairing a capable model with a strict checker, applies well beyond mathematics.

Biology: a new enzyme system with CRISPR-like repeats

On September 23, Anthropic revealed that it runs its own biology lab in the Bay Area, where Claude proposes hypotheses and human scientists run the experiments. The lab works only at biosafety levels 1 and 2 and handles no human pathogens.

Its first public result came from a large automated search. About 950 Claude agents spent 21 hours and 210 million tokens combing genome databases, collected more than 200,000 reverse transcriptases (enzymes that copy RNA into DNA), narrowed 3,500 new candidate systems down to the 20 most promising, and wrote up each in a human-readable report. One pattern stood out: an enzyme, a partner gene beside it, and a long array of evenly spaced DNA repeats, found mainly in bacteriophages, the viruses that infect bacteria. Anthropic named it array-associated reverse transcriptases, or ART.

The layout resembles a CRISPR array, the feature that makes CRISPR gene editing programmable. Lab work has confirmed the repeating structure and that the arrays are expressed as short RNAs. What ART actually does is unknown. CRISPR pioneer Feng Zhang called it an exciting example of AI-assisted discovery, while other scientists quoted by Bloomberg urged caution until experiments show its function.

Faster biology software

A week earlier, on September 17, Anthropic reported that Claude had optimized more than 30 open-source biomolecular models in under four weeks, making them about four times faster on average. It released 36 optimization kits covering structure prediction, binder and sequence design, genomics, and protein language models, plus new GPU kernels and a low-memory mode that models systems of over 10,000 residues and atoms on a single GPU node. Two staff scientists supervised the work without prior experience in GPU engineering. Anthropic also opened a protein design competition with Adaptyv Bio.

The same day, Anthropic launched a Life Sciences Verification Program that loosens biology safeguards for vetted research organizations. Our Fable 5.1 and Mythos 5.1 guide explains how its grants work.

Physics: a nine-loop scattering amplitude

Particle physicists compute scattering amplitudes, the probabilities of particle collisions, in increasingly precise layers called loops. For one benchmark quantity in a simplified theory of particle physics, the six-particle amplitude in planar N=4 super-Yang-Mills theory, the field had reached eight loops, and nine was considered possibly out of reach by direct calculation.

Anthropic's September 25 post describes how Claude Fable 5.1 computed it at nine loops over several days, using two independent methods and about 96 CPUs for a week, at a total compute cost of roughly $1,000 to $2,000. The researchers' main instruction was to keep working until told to stop. SLAC physicist Lance Dixon checked the result independently, and a group at the Chinese Academy of Sciences computed most of it separately with AI help within days.

Public health: faster Ebola situation reports

Not every result was a discovery. In a September 22 feature, Anthropic described how a partnership led by the Coalition for Epidemic Preparedness Innovations, the WHO Regional Office for Africa, and the DRC's national biomedical research institute uses Claude in the response to an Ebola outbreak in the east of the Democratic Republic of Congo.

One team built a Claude skill for the daily situation report. It pulls case and lab numbers from each health zone's slide deck, checks them against the previous day, flags and explains changes in trends, and summarizes district reports. Data teams also use Claude to compare disease models they previously had no time to try. This is the kind of recurring, file-heavy reporting that Claude Cowork's scheduled tasks and skills are built for.

Codebreaking: a disputed historical cipher

In September, the benchmark company Vals AI reported that Claude Fable solved a cipher printed in a 17th-century text, in 44 minutes. Security expert Bruce Schneier shared the claim but wrote that he was not sure the result was correct, and readers pointed to a published refutation and questions about whether the cipher appears in every historical copy of the text.

Treat this one as unconfirmed. It contrasts with Anthropic's own July 2026 report that Mythos Preview found previously unknown weaknesses in modern cryptographic algorithms, including the post-quantum candidate HAWK, which cryptographers could check directly.

What these results have in common

• Many agents, one shared plan. The Fermat proof and the enzyme search both split work across dozens or hundreds of agents coordinated through a shared structure, a theorem graph or a candidate list.
• Long, mostly unattended runs. Days of work with occasional human direction, not a single prompt and answer.
• Built-in verification. Lean checks every proof step, physicists recheck amplitudes, and lab experiments test biological hypotheses. The disputed cipher is the one result without a clean independent check.
• Humans set the goal and judge the outcome. In every case people chose the problem, steered occasionally, and decided whether the result was real.

Anthropic's October 1 essay, "Claude-shaped science," argues that this pattern, rather than any single result, is the change: work that is broad, tedious, and checkable is becoming cheap enough to try.

What this means if you use Claude at work

The models behind these results are the ones in your model picker. Fable 5.1 and Opus 5.5 run in Claude Cowork, and the working style that made these projects succeed carries over to ordinary office tasks:

1. Give the whole job and a definition of done. "Compile last quarter's churn by segment into a two-page memo with a table" works better than a series of small prompts.
2. Build in a check. Ask Claude to reconcile totals against the source file, or to list every number it could not verify.
3. Let long tasks run. In the merged Claude app, tasks continue in the cloud while you close your laptop. Our merge explainer covers how that works.
4. Turn repeats into skills. The Ebola reporting skill is a good template for any daily report built from several files.
5. Keep a human verdict. Review the result before it leaves your team, especially numbers and claims.

If you are deciding which model to use for heavier work, our Opus 5.5 and Sonnet 5.5 comparison explains the trade-offs between capability, speed, and usage.

Frequently asked questions