An old factoring challenge became a perfect AI headline. The maths was familiar, the security scare was overstated, and the part investors should care about is how much expert work one person can now direct.
An 862-bit integer is almost perfect venture-capital theatre. It carries the RSA name, sounds dangerous to anyone who remembers that RSA secures things, and has an answer that can be checked in seconds once someone produces the factors. On September 3, Cognition researcher Eric Lu posted a 130-digit number and two words: "divides RSA-260." It did.
The timing made the story irresistible. Two days earlier, Bloomberg had reported that Cognition was set to raise around $1 billion at a $47 billion valuation. Five days after the factor appeared, Cognition announced more than $2 billion of financing at a $48 billion valuation. Lu's technical account arrived on September 9, six days after the post that set the story loose.
Read that sequence too quickly and the joke writes itself: a $48 billion AI company bought itself a cryptographic legend for $400,000.
The joke is wrong in an important way. Lu had started the work weeks before the funding report, and the RSA-260 run used spare or fragmented cluster capacity that Cognition says had no marginal cost. There is no public evidence that the factorisation was staged around the financing. The chronology is still useful, though, because it shows how easily a messy human-machine project can be flattened into a clean story about machine intelligence.
That story was "AI cracked RSA." Lu's account is more interesting. He used the General Number Field Sieve, the same family of methods behind previous public factoring records, and pushed a heavily modified CADO-NFS implementation onto GPUs. His technical summary is unusually plain: "essentially no algorithmic advancements", only "good old performance engineering". The mathematics was established. The change was in how cheaply, and with how little human staffing, it could be made to run.
The project began with spare GPUs
The origin story is almost comically prosaic. Cognition was trying to improve the scheduler for its AI clusters. Large language-model workloads consume machines in groups, which leaves odd nodes stranded inside racks. Lu wanted useful jobs that could soak up those scraps of compute and disappear instantly when something more important needed the hardware.
Lattice sieving happens to fit that problem beautifully. It can be split across billions of small work units, it does not care much which single node handles a particular unit, and it can be pre-empted. The expensive stage of a famous cryptographic factoring problem was, from a scheduler's point of view, excellent filler work.
At 12:11 a.m. Pacific time on August 13, Lu pointed Devin at a practical engineering task: build a fast GPU drop-in replacement for CADO-NFS's CPU lattice siever. Two hours later he added that it should handle the parameters used for RSA-250. Then he went to bed. Seven hours after that, according to his account, Devin had a working version.
That little sequence tells us more about the state of coding agents than most benchmarks do.
Over the following weeks Lu kept pushing the system across the rest of the pipeline. Polynomial selection moved onto GPUs. The lattice siever was rewritten. The CADO head was changed to cope with the volume of work. Filtering and merge stages were parallelised. The block Wiedemann linear-algebra stage was moved onto GPUs. Even the square-root step needed surgery after an integer-size overflow derailed the final stretch.
The result was a record. RSA-260 has 260 decimal digits, or 862 bits, and is now the largest RSA Factoring Challenge number publicly factored with a general-purpose method. The previous record, RSA-250, required roughly 2,700 CPU core-years in 2020. Lu estimates the RSA-260 run consumed about 4,900 GPU-days, or 13.5 GPU-years. At market rates his headline estimate is about $400,000; the appendix gives $414,000. Cognition says its GPU lattice siever reduced cost by roughly tenfold against the previous public state of the art.
That is substantial engineering. Dismissing it as mere implementation would be as lazy as declaring that AI had broken cryptography. Old mathematics made cheaper can become economically new. Nobody needs a new theorem if a known method suddenly costs one-tenth as much to deploy.
The more provocative number is not $414,000.
It is 3,328.
Three thousand three hundred and twenty-eight messages
Lu says he sent 82,702 words in 3,328 messages across 192 of the 233 Devin sessions used in the project. The agents also started 101 child sessions themselves; 36 of those received no direct intervention from him. That is a strange way to describe either an autonomous mathematician or a glorified autocomplete.
Lu's description of his role sounds more like a technical lead running an unusually tireless team. He set priorities and scope. He stopped useless measurements. He pointed out directions the agents had missed, imposed comparable experimental frameworks and caught them when they abandoned promising work too early. The system for keeping measurements and benchmarks coherent, he admits, required his hand on the tiller.
CADO-NFS mattered for much the same reason. It already broke the factoring problem into stages with stable interfaces and reference implementations. Lu says that human-designed decomposition was essential, and that the agents became more confused as their code moved further away from upstream CADO-NFS. The agents were strongest inside a structure that humans had spent years making legible.
None of this reduces Devin to autocomplete. In Cognition's account, the agents handled measurements, cluster operations, debugging and large amounts of optimisation work. Lu says the project took about three weeks from first prompt to factors and replaced what would probably have been months of work by a team of specialists. That comparison comes from Cognition and has not been independently audited, but the division of labour described in the write-up is revealing.
Lu did not disappear from the loop. His job moved upward. Instead of writing every kernel and operating every job himself, he spent more of his time deciding what the swarm should do, whether its results could be trusted and when to change course.
That is a much larger commercial proposition than "AI can code". One strong engineer may be able to supervise an amount of implementation that used to justify several salaries. If that pattern survives contact with ordinary corporate work, it changes the economics of software engineering, technical consulting and parts of scientific research.
It also creates a delicious problem of authorship. The factorisation rests on decades of mathematics, CADO-NFS, GPU hardware, Cognition's infrastructure, whatever models were operating underneath Devin, and Lu's judgment. Lu himself says he cannot cleanly apportion credit among the participants. Marketing has fewer philosophical difficulties. The achievement gets attached to the product.
The easy "wrapper company" insult does not quite survive contact with the facts either. Cognition post-trains models rather than merely exposing somebody else's API. SWE-1.7 was trained from a Kimi K2.7 base, and on September 10 Cognition disclosed that its newer SWE-2 was post-trained from Kimi K3. That still leaves a useful question about where the underlying capability originates, but it makes the company's bet clearer: base-model intelligence can be sourced while value is built in post-training, orchestration, tools, memory, permissions, evaluation and the customer relationship.
That is the investment question hiding inside the maths. Cognition is neither a magical laboratory whose agent woke up and became a number theorist nor an empty shell around another company's model. Its valuation ultimately rests on whether it can repeatedly turn sourced intelligence into more completed work per human supervisor.
The cryptography scare is the wrong scare
RSA-260 is a wonderful headline because "RSA" is also the name of a cryptosystem people use in the world. The association did a lot of promotional work that the number itself does not justify.
RSA-260 is an 862-bit challenge number. Under current NIST guidance, approved RSA signature generation uses moduli of at least 2048 bits. Lu's own scaling table estimates that RSA-2048 would require about 91.2 billion times the computation of RSA-260 under the same classical GNFS assumptions. His estimated GPU bill is $3.77 × 10^16. Whatever September 3 proved, it was not that today's 2048-bit RSA keys are suddenly waiting to be harvested by a startup with a GPU account.
RSA-1024 is more uncomfortable. Cognition's table puts it at 77.9 times the computation of RSA-260 and roughly $32.3 million in GPU cost. Lu's text rounds that to about $30 million and says moderate further engineering might cut the number again. RSA-1024 has been considered insecure for years, so this is not a new cryptographic emergency. It is a repricing of an old one.
And repricing matters.
Security is economic long before it becomes absolute. A system can remain mathematically difficult while becoming affordable to a larger class of attacker. GPUs remove the need for exotic custom hardware. Agent-assisted engineering can reduce some of the specialist labour needed to make those GPUs useful. The population capable of attempting an attack grows before the mathematics changes at all.
That is where RSA-260 escapes the cryptography story.
Consider the ingredients. A mature body of theory. Good open-source software. A problem that can be divided into testable pieces. Progress that can be measured. Compute that can be rented or, in Cognition's case, scavenged from capacity that would otherwise sit idle. Under those conditions, a highly capable person can keep feeding implementation problems to agents, inspect what comes back and decide where the next hour of machine effort should go.
A surprising amount of technical work has that shape.
As implementation gets cheaper, more of the bottleneck moves into judgment: selecting the problem, designing the benchmark, spotting a bogus result and knowing when a local optimisation has quietly broken the system somewhere else.
That is what a $48 billion valuation should eventually be forced to answer for. Cognition's demonstration is impressive if one researcher compressed months of specialist work into three weeks. But the next evidence cannot be another spectacular one-off. It has to be boring enough to audit: interventions per completed project, rework, failure rates, supervisor skill, performance outside familiar codebases and output per human hour.
If an ordinarily good engineer can reproduce Lu's multiplication of effort, Cognition has evidence for something much larger than a coding assistant. If the effect depends on an exceptional operator correcting thousands of subtle mistakes, the economics look very different.
RSA's old challenge numbers were created to put a price on factoring. Cognition has accidentally made RSA-260 useful for measuring something else.
The number to watch now is the price of organising expertise.