AI has achieved a perfect score at IMO 2026. In July 2026, Huawei’s Celia and Xiaohongshu’s dots-note 3.0 each posted an officially graded 42 out of 42 at this year’s International Mathematical Olympiad, delivering the flawless result that traders on Manifold had been pricing as increasingly likely for weeks; TechXplore and DigitalApplied were among the first outlets to report Celia’s run, while the South China Morning Post confirmed dots-note 3.0’s score as the first AI perfect score in the contest’s history. The play-money market pushed its odds on this exact outcome into a 94-96 percent range by late July, with 513 unique traders on the book, up from roughly 91 percent earlier in the month, as the competition played out in Shanghai. The result builds directly on 2025, when frontier models first reached gold-medal level but stopped short of a perfect run, and Tech Insider’s own tracking of the contest corroborated both scores as they were confirmed. Independent tallies went further still: outlet BIT.FAN counted three systems, including Anthropic’s Claude Fable 5, also landing a 42/42 on the same problem set, albeit outside the IMO’s official grading process.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
The market everyone is suddenly betting on
The question was blunt: “Perfect score achieved by an AI model in the International Math Olympiad (IMO) 2026?” It traded on Manifold, a play-money forecasting platform, and became one of the more active contracts on the site through July 2026. By July 19, 2026, the market priced YES at about 94 percent implied probability with 513 unique bettors, up from roughly 91 percent earlier in the month. Other snapshots from the same window put the figure as high as 96 percent. Weeks later, two systems delivered the exact outcome traders had been pricing.
A perfect score means an AI system solves all six IMO problems and receives full marks, 7 points each, for a total of 42 out of 42, exactly what Huawei’s Celia and Xiaohongshu’s dots-note 3.0 each achieved in July 2026. That is a much higher bar than the gold medal AI systems reached in 2025. Manifold’s resolution follows the same framework as the site’s companion “AI wins gold” market, but requires the top result rather than the medal threshold. If you are new to how these contracts convert crowd sentiment into a number, our explainer on how prediction markets work walks through the mechanics.
What the 94-96 percent odds actually implied
An implied probability in the 94-96 percent range told you the crowd saw a perfect AI score as close to a near-certainty, but not a lock. Odds ran from roughly 16-to-1 in favor of YES at the low end of that range up to about 24-to-1 at the high end. That was aggressive given the specific difficulty involved, and for weeks it was worth separating the forecast from the fact, since no AI perfect score had yet been officially confirmed at IMO 2026. By late July 2026, the forecast became fact: Huawei’s Celia and Xiaohongshu’s dots-note 3.0 both delivered the officially graded 42/42 the market had been pricing. Our primer on implied probability covers why odds in this range still leave room for a surprise, and, in this case, why the surprise never arrived.
Why the odds kept climbing in July
Two things moved the market. First, the 67th IMO took place in Shanghai, at Shanghai High School, across a program that ran from July 10 to 21, 2026, with the two proof-based contest papers written mid-event. With the exams complete and results confirmed later that month, traders had correctly priced in the widely held expectation that frontier labs would run their models on this year’s problems, as they did in 2025. This time, two of them, Huawei’s Celia and Xiaohongshu’s dots-note 3.0, posted a flawless 42/42.
Second, the capability trend had been steep. Earlier in July the same market sat around 91 percent. The move into the 94-96 percent range by late July tracked a year of rapid progress, from no AI gold before 2025 to multiple gold-tier systems in 2025 and, within months, a confirmed perfect score in 2026. That broader momentum showed up outside the IMO too: OpenAI’s GPT-5.4 scored roughly 95 percent on the 2026 USA Mathematical Olympiad, a separate competition-math benchmark, reinforcing that the gains behind the IMO result weren’t an isolated spike. For readers following the broader model race, our comparison of the leading 2026 frontier models gives the wider context.
IMO 2026 Results: Two AI Systems Hit a Perfect 42/42
The wait ended in July 2026. Huawei’s Celia and Xiaohongshu’s dots-note 3.0 were both reported to achieve a perfect 42/42 through the IMO’s own formal grading process, effectively a 100 percent score across the exam’s six problems, each worth seven points, as TechXplore laid out in its July 2026 write-up. One report described it as the first flawless AI score under the competition’s official process. XenoSpectrum‘s IMO 2026 coverage put the field at 666 human competitors, of whom only 7 matched the AI systems with a perfect 42/42 of their own, underscoring just how rare a flawless run remains even among elite human solvers. It is the outcome the Manifold market had been pricing at 94-96 percent for weeks, arriving after the July 10-21 contest window closed.
Xiaohongshu’s result carries an added distinction: dots-note 3.0 is described as the first AI system in IMO history to combine a full score with a gold-medal-level result, and only the second AI model overall, after Google DeepMind’s Gemini Deep Think in 2025, to receive an officially graded gold-level result. Industry tracker APIMart broke the run down problem by problem, confirming dots-note 3.0 covered all six IMO 2026 problems for the full 42/42 tally. That matters because, as covered below, OpenAI’s 2025 score was graded independently by former medalists rather than through the IMO’s own coordinators. That distinction separates officially certified gold from a strong result graded outside the formal process. Huawei’s Celia reached the same 42/42 mark, rounding out the first year a perfect score has been recorded at the IMO by machine.
How AI reached gold in the first place
The story starts at IMO 2025 in Australia. Google DeepMind entered an advanced version of Gemini with Deep Think, which solved five of the six problems and scored 35 out of 42, the gold-medal threshold. DeepMind said its system was part of an inaugural cohort graded and certified by IMO coordinators using the same criteria applied to student solutions, and that it worked end-to-end in natural language within the standard 4.5-hour time limit per paper.
OpenAI reported an identical outcome in the same window: an experimental reasoning model also scored 35 out of 42 by solving the same five problems. OpenAI’s solutions were graded independently by three former IMO medalists rather than by the official coordinators. Both systems missed problem 6, the hardest on the paper. Only six of the roughly 630 human contestants earned full marks on that problem, which frames how big a jump from 35 to 42 really was. It’s a gap that Huawei and Xiaohongshu closed within the year.
DeepSeek and the open-source catch-up
The next escalation came from open weights. On November 28, 2025, DeepSeek released DeepSeek-Math-V2, described as the first open model to reach IMO gold-medal level. The system, a large mixture-of-experts model published on Hugging Face and GitHub, solved five of six IMO 2025 problems for a gold-tier score and posted 118 out of 120 on a Putnam exam in extended-compute testing. Its design leans on a generator-verifier loop that checks each proof step and feeds errors back for self-correction, an approach aimed at valid reasoning rather than just final answers. The technical report is on arXiv.
The significance for markets is that gold-level math reasoning was no longer locked behind a single lab’s API. That broadened the field of systems that could, in principle, produce a perfect run in 2026, which is part of why traders had grown more confident. By July 2026, that’s exactly what happened: Huawei and Xiaohongshu each reached a perfect 42/42.
The gap between gold and a perfect 42/42
This was where the 94-96 percent odds got interesting. A gold medal at the IMO does not require solving every problem. A perfect score does. In 2025 the best AI systems, and the strongest open model, all stalled on the same sixth problem that stumped nearly every human in the room. Combinatorics and the hardest geometry problems had been the areas where automated proof systems were least reliable, and a single unsolved problem was the difference between the medal traders already expected and the flawless result the market demanded. The crowd bet that another year of scaling and better verification would close that gap. In July 2026, it did. Huawei’s Celia and Xiaohongshu’s dots-note 3.0 each solved all six problems, including the combinatorics and geometry problems that had previously proven hardest to automate, for the first officially graded 42/42 scores in IMO history.
A timeline of AI at the Math Olympiad
| Date | System | Result | Score |
|---|---|---|---|
| July 2024 | Google DeepMind AlphaProof / AlphaGeometry | Silver-medal level | 28/42 |
| July 2025 | Gemini Deep Think (Google DeepMind) | First officially graded AI gold | 35/42 |
| July 2025 | OpenAI experimental reasoning model | Gold level, graded by former medalists | 35/42 |
| Nov 28, 2025 | DeepSeek-Math-V2 | First open-weight gold-level model | Gold tier (5/6 solved) |
| July 2026 | Celia (Huawei) | First officially graded AI perfect score | 42/42 |
| July 2026 | dots-note 3.0 (Xiaohongshu) | First AI to pair a perfect score with a gold-medal result (2nd officially graded gold overall) | 42/42 |
Who actually delivered at IMO 2026
Going into the competition, the obvious candidates to produce a perfect run were the same labs that reached gold in 2025: Google DeepMind, OpenAI, and, on the open-weight side, DeepSeek. Instead, two entrants outside that group delivered the result first. Huawei’s Celia and Xiaohongshu’s dots-note 3.0 both posted an officially graded 42/42 in July 2026, with dots-note 3.0 becoming only the second AI model, after Gemini Deep Think, to receive that specific certified-gold distinction. Outside the official grading process, the field widened even further: an independent July 2026 benchmark run by AllHereAI found Claude Fable 5, GPT-5.6 Sol, Kimi K3, and AxiomProver all reaching 42/42 on the same problem set, a result AI researcher Deedy Das corroborated on X after running his own tests on those same four systems that month, and one that outlet BIT.FAN echoed separately, counting three models including Claude Fable 5 at a 42/42 score and verifying AxiomProver’s perfect run on its own. Industry outlet 36Kr Europe put numbers on how those runs were produced: Claude Fable 5 reached its perfect 42 in about 2.5 hours of compute for roughly $51, GPT-5.6 Sol (xhigh) took 3.8 hours and about $20, and Kimi K3 needed 17.4 hours and roughly $31. OpenAI’s own July 2026 release notes for GPT-5.6 showed the model scoring 92.2 percent on the BrowseComp benchmark, part of the broader capability jump behind that IMO run. Independent benchmark tracker BenchLM.ai also listed Anthropic’s Claude Opus 5 at 42/42 on its public IMO 2026 leaderboard, which by September 2026 showed the model leading with an aggregate score of 82.72, though none of these unofficial runs went through the IMO’s own certified grading the way Celia’s and dots-note 3.0’s scores did. It’s a reminder of how quickly the field of labs capable of frontier-level math reasoning has widened. For a related contract on machine mathematical reasoning, see our coverage of the market on AI solving an open math conjecture.
Why a tech and finance reader should care
Beyond the novelty, this market was a live gauge of how fast the crowd thought automated reasoning was improving. A confirmed perfect score is now a clean, verifiable benchmark that AI can produce rigorous, fully correct proofs at the ceiling of human competition math, which feeds directly into debates about model capability, research automation, and the valuations attached to the labs building these systems. Prediction markets have become a fast, if imperfect, sentiment layer on top of that story. If you want to compare how the major venues operate, our Kalshi versus Polymarket breakdown is a useful starting point, and a related Polymarket contract on whether AI wins IMO gold in 2026 tracked a lower bar that this Manifold outcome has already cleared.
Frequently asked questions
Has an AI actually scored 42/42 at IMO 2026? Yes. In July 2026, Huawei’s Celia and Xiaohongshu’s dots-note 3.0 each achieved an officially graded 42/42, the first perfect AI scores recorded at the IMO. Manifold’s odds had already reached the 94-96 percent range in the weeks before the confirmation.
Which AI systems scored a perfect 42/42 at IMO 2026? Huawei’s Celia and Xiaohongshu’s dots-note 3.0, both graded through the IMO’s official process in July 2026. dots-note 3.0 is also the second AI system, after Gemini Deep Think, to receive an officially certified gold-level result. Outside that official process, independent July 2026 testing by AllHereAI, researcher Deedy Das, and outlet BIT.FAN also put Claude Fable 5, GPT-5.6 Sol, Kimi K3, and AxiomProver at 42/42, and BenchLM.ai’s September 2026 leaderboard listed Claude Opus 5 at 42/42 with an aggregate score of 82.72, though none of these carry the IMO’s own certification.
What is the difference between the gold score and a perfect score? Gold at the IMO required about 35 of 42 points in 2025 and did not need all six problems solved. A perfect score requires full marks on every problem, including the hardest, which no AI managed in 2025. Huawei’s Celia and Xiaohongshu’s dots-note 3.0 became the first to do it, in July 2026.
Which AI models reached gold in 2025? Google DeepMind’s Gemini Deep Think and an OpenAI experimental reasoning model each scored 35/42. DeepSeek-Math-V2 later reached the same gold level as an open-weight model in November 2025.
Where is IMO 2026 being held? In Shanghai, China, at Shanghai High School, across a program running July 10 to 21, 2026.
Is this a bet I can make with real money? Manifold runs on play-money by default. This article is not betting or investment advice, and availability of any real-money venue depends on your jurisdiction.
The Bottom Line
Sources
- Manifold — Perfect score achieved by an AI model in the International Math Olympiad (IMO) 2026?
- Google DeepMind — Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the IMO
- CBS News — Humans triumph over AI at annual math Olympiad, but the machines are catching up
- South China Morning Post — DeepSeek releases first open AI model with gold-level scores at maths olympiad
- arXiv — DeepSeekMath-V2: Towards Self-Verifiable Mathematical Reasoning
- IMO Official — IMO 2026 edition details
- Polymarket — AI wins IMO gold medal in 2026
Prediction markets carry risk and are not investment or betting advice. Market prices reflect crowd sentiment and can be wrong. Availability is restricted by jurisdiction: Polymarket is not available to US persons, and Kalshi is a CFTC-regulated US exchange; Manifold operates primarily on play-money. Participation is limited to those 18 or older, or 21 or older where applicable. If gambling is a problem for you or someone you know, call 1-800-GAMBLER for help.
