Skip to content
← All news

News3 min read

30/30 IPhO, 6/6 IOI, 100/100 IBO

IPhO 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.

Since July, the same open harness has finished IPhO and the quantum run we reported then, and has been pointed at three more olympiads. Each one reached the top of its own scale. None of these scores came from an official jury, and the scales are different, so each comes with a note on what was graded and by whom. The notes matter as much as the numbers.

Five exams · what was scored, and who scored it
IPhO 2026 · theory30.00 / 30.00, both models100%our audit against the official marking scheme
IOI 20266 / 6 problems at 100%100%Codeforces
IBO 2024 · theory100 / 100 tasks; 400 / 400 verdicts100%the official answer key
IChO 2026 · theory68 / 68 subquestions formalized, both models100%Lean, plus our own review gates
QIT40 / 40 tasks proved in Lean100%Lean, plus automated semantic review

How a run stays blind ​

Every run here starts the same way: the workers get the problem and nothing else. A builder writes, a reviewer objects, round after round. The official solutions, marking schemes and answer keys come in only after the run is over, for grading.

Answer-blind · then graded
ProblemBuilderReviewerOfficial keyGraderafter the run

Swipe the diagram sideways to see all of it →

1/0 The workers get the problem, and nothing else.

IPhO 2026: 30 of 30 on theory ​

Both the GPT-5.6 Sol and the Kimi K3 Max runs answered all 23 theory subparts and scored 30.00 / 30.00 (GPT, Kimi). The workers never saw the official solutions or marking schemes. Those came in only after the runs had finished, for grading.

The caveat is the grader. These are our own estimates against the official solutions and itemized marking schemes, not a jury's adjudication.

The Lean side went further than in July. The formal release splits the paper more finely, into 41 parts, and every one of them now has a Lean formalization. Formalizing is harder than answering, and the two models separate here: the GPT-5.6 Sol formalizations are estimated at 47.45 / 50, Kimi's at 29.90 / 50. Every proof elaborates without sorry, admit or custom axioms.

IPhO 2026 · share of the points, our audit
Natural-language answersof 30 points
100.0%
100.0%
Lean formalizationsof 50 points
59.8%
94.9%

IChO 2026: 68 of 68 formalized ​

All 68 numbered theory subquestions across the nine problems now have Lean formalizations that passed both formalization review and proof review, for GPT-5.6 Sol and for Kimi-K3 (32 answer-blind plus 36 more). The comparison that matters is the same models without the review loop. Run as a plain Codex /goal on the same 68 targets, they had 32 / 68 and 31 / 68 accepted.

IChO 2026 · formalizations accepted, of 68
32
68+36
GPT-5.6 Sol
31
68+37
Kimi-K3
The same models on the same 68 targets, with and without the review loop. The small pink numbers are the loop's gain.

The fine print:

  • 68 / 68 is formalization coverage, not answer accuracy. Lean checks deductions from the encoded inputs. It does not certify that the encoding matches the official chemistry.
  • Not every target came from the original inputs alone. GPT has 66 targets from the original inputs and 2 conditional ones. One of the conditional targets, T8-A8, uses two disclosed contest-model axioms that we authorized.
  • Kimi's 68 is two campaigns. One is 32 answer-blind targets; the other formalizes 36 natural-language answers Kimi had already written. Three of those 36 are conditional and two carry authorized answer corrections.
  • The answer scores are generous and unofficial. Graded against the official key with credit for equivalent forms and reasonable rounding, GPT scores 424.5 / 437 and Kimi 417.5 / 437, against 415 and 340.2 for the plain /goal runs.

IOI 2026: six of six ​

All six problems over both days passed at 100% when the final submissions were judged on Codeforces.

IOI 2026 · final submissions, judged on Codeforces
1Ball Machine100%
1Monuments100%
1Tiling Game100%
2Classroom Game100%
2Magic City100%
2Partition100%

The annotated solutions state the guarantee each one meets. What you can check yourself is narrower: the repository's verify.sh compiles all six and replays every released Day 1 example. Codeforces' hidden tests remain the external check, and the verdicts are not linked publicly.

IBO 2024: 100 of 100 theory tasks ​

Theory tasks

100/100

50 in Theory A and 50 in Theory B.

True/false verdicts

400/400

Every one matches the official answer key.

Extraction errors

0

No answer lost in formatting.

All 100 Theory A and B tasks were solved, and all 400 true/false verdicts match the official answer key. Two limits apply. The practical exams are excluded entirely, and this is agreement with an answer key, not an official points total. It is also the 2024 paper, because IBO withholds each paper for two years.A paper that has been public that long may appear in a model's training data.

Quantum information theory: 40 of 40 ​

In July, the blind QIT run stood at 37 of 40 end to end. After repair it is 40 / 40, alongside 36 / 36 on quantum algorithms (QAlg).

QIT · end to end

40/40

Up from 37 of 40 in July.

QAlg

36/36

Unchanged since July: every task.

Every task compiles, has a complete proof with no sorry or admit, and passes a full lake build. The caveat is the one we drew in July: the semantic review, which judges whether a formal statement says what the problem says, is automated, not an independent human audit. Both sets rebuild with one script.

IPhO 2026 · IChO 2026 · IOI 2026 · IBO 2024 · 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 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 News#1 on Lean-Eval: 172 proofsThree weeks after second place, a fully agentic GPT-5.6 run is first on the Lean-Eval leaderboard, with 172 research-level problems solved. Every accepted proof is sorry-free and independently re-verified.Read it