AI Scored 42/42 at the IMO. Who Marked It?

Several AI systems now say they scored full marks on the International Mathematical Olympiad (IMO) 2026. A perfect score sounds like one fact. When I checked the sources, it turned out to be at least three different kinds of claim, and the difference is who marked the paper.

Transcript

AI just got full marks on the hardest school maths exam in the world. But who actually marked it? The exam is the International Mathematical Olympiad. Six problems, seven marks each, forty-two in total. This year, in Shanghai, six hundred and sixty-six students sat it. Only seven got full marks. Two AI systems are reported to have scored forty-two out of forty-two under the organisers' own marking. Huawei's Celia, and RedNote's dots note three. Those reports come from the companies. I couldn't find a marking report from the Olympiad itself. Then there's a second kind of test. An investor, Deedy Das, ran Claude Fable five, GPT five point six Sol, and Kimi K three himself. All three ended on forty-two. But the markers were AI agents, and the page itself says to treat the scores as strong, not authoritative. Look at the first attempts. Fable got forty-two straight away. Sol scored thirty-nine, then reached forty-two after repair rounds. Kimi scored thirty-six, then forty-two. Same final score. Different journeys. Settings matter too. The same model, Sol, scored twenty-eight on default, forty-two on extra high, and thirty on max. One setting moved the score by fourteen marks. Open models are in the race. NVIDIA says Nemotron three Ultra scored thirty out of forty-two, above the gold line of twenty-nine. So when you see a perfect score, ask four things. Who marked it? How many attempts did it take? What settings were used? And what did it cost? One run was about fifty-one dollars. Another, about twenty. A perfect score is a headline. The marking is the evidence. Follow for sources, not hype.

The exam

The IMO is the hardest school maths competition in the world. There are six problems worth seven marks each, so 42 is a perfect score. In 2026 it was held in Shanghai on 15 and 16 July. The dots team's page reports 666 contestants from 117 countries, and only seven human contestants scored 42.

Claim 1: marked by the organisers (company-reported)

Two systems are reported to have scored 42/42 under the organisers' own marking:

• dots-note-3.0 from Xiaohongshu (RedNote). The team's page says the result followed official marking organised by the IMO committee, and that the model received the original LaTeX statements with no human rewriting.
• Celia from Huawei, reported as officially graded by Tech Insider. Taipei Times attributes the result to the companies' own statements.

Both claims come from the companies or from coverage of them. I did not find a marking report published by the IMO itself, so I treat this as reported, not independently confirmed.

Claim 2: marked by AI agents (self-run)

Investor Deedy Das ran several models against the problems himself and published every solution in a public repository. Three ended on 42/42:

Model	First pass	Final	Total time	Total cost
Claude Fable 5	42	42	2.5 h	$51.05
GPT-5.6 Sol (xhigh)	39	42	3.8 h	$20.54
Kimi K3	36	42	17.4 h	$31.40

Two details matter. First, the markers were Claude-based verifier agents, and the README says: "Graders are Claude-based agents, not human medalists; treat scores as strong but not authoritative." Second, only Claude Fable 5 reached 42 without a repair round. For the other two, the reviewer's specific defect findings were shown to the model (never other solutions) and it was given more sessions, then the result was re-verified. Costs are the repository's totals, which include rounds lost to outages; its "clean" figures are lower.

One setting moved the score by 14 marks

The same repository ran GPT-5.6 Sol at several effort settings:

Setting	Final score
Default	28
Extra high (xhigh)	42
Max	30

Same model, same problems. Which setting you pick can matter more than which model you pick, and "max" was not the best here.

An open model, with a gold line

NVIDIA says its open-weight Nemotron 3 Ultra scored 30 out of 42, above the gold line of 29. That is NVIDIA's own report, so it belongs with the company-reported results.

Four questions to ask of any perfect score

1. Who marked it? The organisers, human experts, or another AI?
2. How many attempts did it take? First pass, or after repair rounds?
3. What settings were used? Effort, tools, time limit.
4. What did it cost? One run was about $51, another about $20.

None of this makes the results less impressive. It makes them easier to compare, which is the point of a benchmark.

What I could not verify

• The IMO's own results page for the AI entries.
• Huawei's primary statement for Celia.
• One blog also lists AxiomProver at 42/42, but it is not in the repository files I read, so I left it out.

Sources

• dots team, IMO 2026 result page: the Xiaohongshu claim, problem delivery and context.
• Deedy Das, imo-2026 repository: scores, first-pass tables, grading method and costs.
• Tech Insider, AI perfect score at IMO 2026 and Taipei Times, AI models score 100 percent: coverage of the Huawei and Xiaohongshu claims.
• NVIDIA, 21 July 2026 post on Nemotron 3 Ultra, summarised at postcutoff.com.
• My own article, AGENTS.md: One Rulebook for Cursor, Claude Code and Codex, is another case of checking claims against primary documents.