Skip to content
← All news

News1 min read

6/6 at IMO 2026, Lean-checked

A fully agentic run solved all six problems of the 2026 International Mathematical Olympiad, every proof machine-checked in Lean 4. Two workers did it independently, the faster in 3.0× less API time than AxiomProver.

All six problems of the 2026 International Mathematical Olympiad, solved by a fully agentic, YOLO-style run with no human in the turn, and every solution formally verified by Lean 4. There is no rubric here and no benefit of the doubt: the kernel accepts the proof or it does not.

Two workers were run independently, and both closed all six.

Both workers

6/6

GPT-5.6 and Kimi-K3 each solved every problem, checked by Lean.

Less API time · GPT-5.6

3.0×

497.7 minutes in total, against 1,496 reported by AxiomProver.

Whole paper · Kimi-K3

$57.70

Less than half the $127.41 of the GPT-5.6 run.

The loop ​

The formal statements are AxiomMath's, released with their own IMO 2026 run. A worker writes a Lean proof against each; Lean checks it, and the AXLE verifier confirms that the proof proves the statement it was given.

Statement → worker ⇄ Lean
StatementWorkerLean 4prooferrors6 OF 6

Swipe the diagram sideways to see all of it →

1/0 Each problem starts from AxiomMath’s formal Lean statement.

The times ​

API time is the time actually spent inside model calls, the one measure that does not reward a slower harness. Here it is against the times AxiomProver reported on the same statements. Compare against AxiomProver, then hover a problem to read the minutes saved or lost.

IMO 2026 · API time per problem, minutes
Q1
24.0 min
87.1 min
38.1 min
Q2
360.0 min
224.3 min
100.4 min
Q3the hardest
869.0 min
343.7 min
187.1 min
Q4
39.0 min
75.6 min
58.7 min
Q5
65.0 min
91.9 min
46.5 min
Q6
139.0 min
212.4 min
66.9 min

The shape of that chart is the point, and it is clearer with one problem per dot. Below the diagonal we were faster; above it, slower.

Our time against AxiomProver’s · log scales
  • Q1 · GPT-5.6: AxiomProver, minutes 24 min, Humanfia, minutes 38 min
  • Q1 · Kimi-K3: AxiomProver, minutes 24 min, Humanfia, minutes 87 min
  • Q4 · GPT-5.6: AxiomProver, minutes 39 min, Humanfia, minutes 59 min
  • Q4 · Kimi-K3: AxiomProver, minutes 39 min, Humanfia, minutes 76 min
  • Q5 · GPT-5.6: AxiomProver, minutes 65 min, Humanfia, minutes 47 min
  • Q5 · Kimi-K3: AxiomProver, minutes 65 min, Humanfia, minutes 92 min
  • Q6 · GPT-5.6: AxiomProver, minutes 139 min, Humanfia, minutes 67 min
  • Q6 · Kimi-K3: AxiomProver, minutes 139 min, Humanfia, minutes 212 min
  • Q2 · GPT-5.6: AxiomProver, minutes 360 min, Humanfia, minutes 100 min
  • Q2 · Kimi-K3: AxiomProver, minutes 360 min, Humanfia, minutes 224 min
  • Q3 · GPT-5.6: AxiomProver, minutes 869 min, Humanfia, minutes 187 min
  • Q3 · Kimi-K3: AxiomProver, minutes 869 min, Humanfia, minutes 344 min

On Q1 and Q4, the two easiest problems, we lose. On Q3, the hardest, GPT-5.6 wins by more than four times. A loop earns its keep exactly where the work is long enough for the loop to matter; on a problem a strong model closes in twenty minutes, the arrangement around it is overhead.

A loop earns its keep exactly where the work is long enough for the loop to matter.

What it cost ​

The open model is the cheaper one by a wide margin.Kimi-K3's figures are its worker and its Codex reviewer combined. Costs are from the release's README, rounded to the cent. Kimi-K3 took longer on every problem and cost less on every problem: $57.70 for the paper against $127.41 for GPT-5.6.

IMO 2026 · cost per problem, US dollars
Q1
8.11
5.73
Q2
34.01
17.76
Q3
60.44
15.55
Q4
8.18
3.69
Q5
5.78
5.18
Q6
10.88
9.80

What is released ​

The formal statements, both sets of Lean solutions and the scripts that reproduce the solving process, pinned to Lean 4.31.0 and Mathlib.

humanfia/hoa-qed/imo2026 · AxiomMath/IMO2026 · HOA

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