Anthropic Engineer Steve Weis Factors RSA 896 With Claude and a Fleet of 2,048 GPUs

On September 19, 2026, Steve Weis, a member of technical staff at Anthropic on the Secure Frameworks team, announced that he had factored RSA-896, one of the RSA Factoring Challenge numbers, using Anthropic's AI model Claude. Weis wrote on September 19 that RSA-896 is a challenge number he factored with Claude on that date. He published the full 270 digit modulus together with the two prime factors, p and q, which means the result can be verified in seconds by multiplying the two numbers and comparing the product to the published modulus.

RSA-896 is an 896-bit number, which works out to 270 decimal digits, and its two prime factors are each 448 bits long, or roughly 135 decimal digits. Weis is not a casual observer of this field. He is a member of technical staff at Anthropic on the Secure Frameworks team and holds a computer science PhD from MIT, where he was advised by Ron Rivest, a co-inventor of RSA. The number he factored belongs to the set of challenge numbers that RSA Security published, in the binary-named 2001 expansion that once carried a $75,000 prize. RSA-896 sits between the 862-bit RSA-260 and the 928-bit RSA-280.

The mathematical method behind the work is not new. Factoring numbers of this size relies on the General Number Field Sieve, an approach developed in the 1990s and implemented by the open source package CADO-NFS. What Weis says is new is the way the work was carried out. He had Claude port CADO-NFS so that it could run on GPUs, the chips that power AI training, and Claude then orchestrated a fleet of those GPUs, drawing on scavenged idle capacity.

The scale of that orchestration is the striking part. According to Weis, the run used up to 2,048 GPUs at once, over about 10 days, for roughly 30 GPU-years of total computing. Spreading 30 GPU-years across 2,048 machines is what let the job finish in ten days rather than three decades. The result landed less than three weeks after a related milestone. On September 9, Eric Lu of Cognition published the factoring of RSA-260, a 260 digit and 862-bit number completed on September 3, using a GPU-accelerated CADO-NFS run managed by a swarm of Devin agents, at about 4,900 GPU-days, or roughly 13.5 GPU-years, and an estimated cost near $400,000.

Key Facts

RSA-896 is an 896-bit semiprime, a product of two primes, with 270 decimal digits. Its two prime factors, p and q, are each 448 bits long, which is about 135 decimal digits each. The full modulus and both factors were published, so anyone can check the arithmetic in seconds by multiplying p by q and comparing the result with the modulus.

OfficeChai reported on September 20 that the run finished on September 19, 2026, and that it used up to 2,048 GPUs at once over about 10 days for roughly 30 GPU-years of total computing. The outlet also said it verified the result itself by multiplying the two published factors and confirming that they give RSA-896.

Weis was explicit about the limits of the result. He said there was no new math, no shortcut and no new factoring algorithm. He also said that ten days and 2,048 GPUs for 896 bits does not translate into cracking the 2,048-bit keys used in practice, and he stated plainly that the work poses no new threat to deployed cryptographic keys.

Crypto Briefing reported on September 19 that the result pushes the public factoring record to 896 bits, beyond the 862-bit RSA-260 that Cognition's Eric Lu publicly solved on September 3, 2026, and beyond the long standing 829-bit RSA-250 record from February 2020. The outlet noted that the number factored was a challenge number rather than a key in production use.

LavX News reported on September 20 that Weis announced he had factored RSA-896 on September 19, 2026, using Claude, describing it as a long standing computational problem in cryptography. LavX also noted that the announcement was terse, with no elaboration on methodology or on Claude's specific role, and that as of its writing the factorization did not appear on the RSA Factoring Records page.

Before these results, the largest publicly factored number of this kind was the 829-bit RSA-250 in 2020, which took about 2,700 CPU-years. The comparison between that effort and the newer runs is stark: 2,700 CPU-years for 829 bits, about 13.5 GPU-years for 862 bits in early September 2026, and roughly 30 GPU-years for 896 bits later the same month. The most common deployed key size, RSA-2048, is about 617 decimal digits and remains far out of reach.

Analysis

The interesting part of this story is not the mathematics. The algorithm used, the General Number Field Sieve, dates to the 1990s, and CADO-NFS is a mature open source implementation of it. Weis said plainly that there is no new math, no shortcut and no new factoring algorithm. What this really means is that the frontier being pushed here is engineering orchestration: taking a tool built for CPUs, getting it to run on GPUs, and then using an AI model to manage thousands of machines at once.

Compare the economics across the recent record runs. RSA-250 in 2020 took about 2,700 CPU-years. RSA-260 in early September 2026 took roughly 13.5 GPU-years and an estimated $400,000. RSA-896 in mid September 2026 took about 30 GPU-years. Each step adds bits, and each step is finished faster in wall-clock terms because the work is spread over more parallel hardware. The growth in cost is real but not explosive, and the wall-clock time is collapsing.

Claude's specific contribution is worth describing carefully. LavX News cautioned that large language models are not dedicated number theory tools, so Claude's contribution likely involved integration with specialized factorization software rather than a novel algorithm. OfficeChai reported that Claude, when asked for a public message about the result, pointed to predecessors and said the credit belongs to the people who built the number field sieve and CADO-NFS over several decades. That is a fair description of where the intellectual work sits.

The bigger picture here is that AI is becoming a force multiplier for mature scientific software, not a replacement for it. The model did not discover a new sieve. It helped port an existing one to new hardware and then ran the logistics. That distinction matters for anyone tempted to read the headline as an algorithmic breakthrough.

Why It Matters

Crypto Briefing argued on September 19 that AI-assisted factoring accelerates cryptographic challenges and should urge a reevaluation of encryption standards, while highlighting the transformative role of AI. The practical framing is narrower than that warning might suggest. The number that was factored was a challenge number, not a key in production use, and 2,048-bit RSA keys, the standard in TLS and much of deployed cryptography, remain far out of reach.

Still, the pace deserves attention. LavX News noted that the cryptographic community has long considered 1,024-bit RSA to be within reach of well resourced adversaries, even though no public result existed as of mid 2026. A result that pushes the public record from 829 bits in 2020 to 862 bits and then 896 bits within the same month shows how quickly AI directed engineering can compress timelines that once stretched across years.

There is also a transparency angle. Because both prime factors were published, the result is independently checkable in seconds, and OfficeChai said it performed that check. At the same time, LavX News observed that the primary post offered no methodology details, and that the factorization did not appear on the RSA Factoring Records page as of its writing. Independent replication and record keeping will determine how much weight the result carries.

Next Up

The immediate next steps are confirmation and documentation. Researchers will check the published factors, and the result may or may not be added to the RSA Factoring Records page, which LavX News noted did not list it at the time of its report. More detail from Weis or from Anthropic about how Claude ported CADO-NFS to GPUs, and how the GPU fleet was orchestrated, would clarify the division of labor between the model and the human.

Beyond that, the obvious question is how far this approach can go. The gap to RSA-2048 is enormous, about 617 decimal digits, and Weis himself said the work poses no new threat to deployed cryptographic keys. But the pattern of using AI to port and orchestrate mature scientific code across large fleets of GPUs is unlikely to stay confined to factoring challenges.