Nobody will read 13 million lines of Lean, and the project does not require it. Divide that by the 129 pages of Wiles' original and you get about 100,800 lines of Lean per printed page [14]. Divide it by the 11 days and you get roughly 1.18 million lines a day [15]. Human review cannot run at that rate. It also does not matter here, because a formalized proof is checked automatically by computer, which is the reason mathematicians want proofs in Lean in the first place [9].
The phrase "AI does mathematics" hides the shape of the run. Those 29,500 intermediate theorems work out to about 441 lines of surviving Lean each [17], at roughly 460 tokens of model output per line that survived [16]. The decomposition is the mechanism itself. Lean is unforgiving in a specific way: arguments build on each other, so one bad line can invalidate everything after it [10]. Cutting the job into 29,500 separately graded pieces is how a failure stays local instead of fatal.
Step selection is the part a team can copy. Prove2Me is open source, it helps agents pick the next move in a long workflow, and Anthropic says it also lowers inference costs [8]. Picking the next move against something that will grade the attempt transfers to any place you already own an oracle: a compiler, a test suite, a schema validator, a reconciliation that has to balance. Compare that with the two other recent results in this vein. Anthropic reported new information about the Riemann zeta function a month earlier [12], and OpenAI used Astra last month to solve several Erdos problems [13]. Those are discovery claims that people have to judge. Formalization is the case where the grader is a program, and that is why the line count is allowed to get silly.
On the evidence: the 11 days, the token count and the theorem count come from Anthropic's own blog post as relayed by SiliconANGLE [1][4], with no independent reproduction reported [19]. Kevin Buzzard, whose work Claude drew on, says autoformalization artefacts are now robust enough to be built upon [11], which is a judgment about usability rather than a replication. There is no compute bill in the account and no count of human interventions behind the phrase "limited high-level input", and those are the two numbers you would need to price a run of your own.
So the tooling question comes down to whether a checker exists for the output and whether a human still has to sign the result, not to how capable the agent is. Draw the 2x2: one axis is whether an automatic checker exists for the output, the other is whether a human still has to sign the result. Checker, no signature required: more agents is close to free, and 13 million lines is a perfectly good answer. Checker plus a human signature: every extra token is review you have bought, and volume works against you. No checker at all: the output lands on your team as a tax, and the demo is quietly asking you to be the oracle. Anthropic's number is impressive because it sits in the first box. Most software teams live in the second, and the useful work is dragging specific tasks out of the third box into the first, one verifier at a time.
