Skip to content
← All news

News1 min read

#2 on Lean-Eval, behind Seed Prover

A general model in a general agent loop is second on Lean-Eval, behind ByteDance's Seed Prover, a system built for formal mathematics. By our count it is 149 problems to 155.

This has since moved

Three weeks later the same loop took first place with 172 problems.

Running on gpt-5.6-sol at max effort, Humanize is second on the Lean-Eval leaderboard. By our count today it has solved 149 of 219 problems. First is ByteDance's Seed Prover, at 155.Those two counts are ours, from the day. The board's published log dates every proof by when it was accepted, and by that log the standings at the end of 29 July were 144 for Seed Prover and 130 for us; it records our 149th on 4 August. The order is the same either way.

Lean-Eval · end of 29 July 2026 · accepted proofs
  1. 1Seed Prover (ByteDance)144
  2. 2HumanfiaGPT-5.6 sol, max effort130
  3. 3Aristotle (Harmonic)124
  4. 4GPT-5.6112
  5. 5Tau (caj.al)104
Counted from the board's own published data (generated 5 October 2026), by acceptance date.

A specialist and a generalist ​

Six problems by our count, fourteen by the board's log: either way it is not a rout, and the gap is worth naming precisely because of what is on each side of it. Seed Prover is a system built for this: a prover, trained and tuned for formal mathematics. What is second is a general-purpose model driving a general-purpose agent loop, with no component anywhere in it that knows Lean specifically.

One prover · one loop
Seed ProverLean proofsHumanizeGPU kernelsKaggle

Swipe the diagram sideways to see all of it →

1/0 Seed Prover is a prover: built, trained and tuned for Lean.

The loop is the same one that runs kernel work and Kaggle competitions.

A specialist should beat a generalist on the specialist's benchmark. That it does so by so little is the interesting number.

That is why we thought the gap was closable.

HOA · the flows

Read next

News#1 on Lean-Eval v1 first solvesLean-Eval froze its first 128-problem release and moved everything older to an archive. On v1 our GPT-5.6 flow has the most first solves, 30, and is one behind the leader on total; on the archive it has 170 of 171.Read it News30/30 IPhO, 6/6 IOI, 100/100 IBOIPhO 2026 theory 30 of 30, IOI 2026 six of six, IBO 2024 theory 100 of 100, IChO 2026 formalized 68 of 68, and quantum information theory 40 of 40. What each score is, who graded it, and what it does not show.Read it News672/672 on PutnamBench, joint #1GPT-5.6 in a worker–reviewer loop proved all 672 PutnamBench statements in Lean, at $44.50 a problem. The top of the official leaderboard is now a four-way tie, and what separates the four is cost.Read it