AI Perfect Score at IMO 2026? Manifold Traders Put It at 77%. A Manifold Markets contract asking whether an AI model will post a flawless 42 out of 42 at the 2026 International Mathematical Olympiad peaked at a 97% implied chance of “Yes” on July 23, 2026, up from 91% ten days earlier, before cooling to 77% as the market’s July 31 resolution date approached. The reversal followed the close of the competition’s exam window in Shanghai on July 15 and 16, official IMO reporting showing just seven human contestants out of 666 scored a perfect 42, and a wave of unverified claims of AI perfect scores from Huawei’s Celia model, Xiaohongshu’s dots-note 3.0, and reports naming GPT-5.6 Sol (xhigh), Claude Fable 5, and Kimi K3. By August 2026, the question edged toward an answer outside the market itself: independent benchmark tracker BenchLM.ai published a public IMO 2026 leaderboard showing Claude Opus 5 posting a flawless 42 out of 42, with Apodex 1.1 trailing well behind at 31 out of 42.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What the Manifold market is asking
The market in question is titled “Perfect score achieved by an AI model in the International Math Olympiad (IMO) 2026?” and trades on Manifold Markets, a play-money forecasting platform where users buy and sell shares in yes-or-no questions. A “perfect score” at the IMO means 42 out of 42 points, earned by fully solving all six proof problems across the two exam days.
According to the market’s resolution notes, a qualifying result must come from an AI model working under conditions close to the human contest: roughly 4.5 hours per exam day, a cap on total compute, and published, verifiable results. The distinction matters. This is a stricter bar than simply matching a gold-medal threshold, because a perfect paper requires solving the two problems that even most human medalists miss.
The odds since July 23, 2026, and what they imply
The Manifold contract peaked at a 97% implied probability of “Yes” on July 23, 2026, according to the market page linked above, up from 91% on July 13. That same page also broke the question down by lab at the time, pricing 96% for OpenAI, 94% for Anthropic, and 77% for Google to be the one that delivers a flawless paper. On July 13, the market had already drawn top-tier 24-hour volume and 473 unique bettors, a sign that the market was drawing real attention rather than a handful of stray trades. But as the July 31 close date has neared, the overall “Yes” price has pulled back to 77% — coincidentally the same level Google’s lab-specific odds held on July 23, though the two numbers track different questions — rather than climbing toward 100%.
An implied probability is not a guarantee. It is the crowd’s aggregate price for a claim, and on a play-money venue it reflects conviction and information flow more than dollars at risk. The pullback from a 97% peak to 77% says traders are now less certain a flawless AI paper will be certified in time, even with several public claims already circulating, a meaningful shift for a benchmark that no AI system reached even at the gold-medal level until 2025. For readers new to the mechanics, our explainer on how prediction markets are pricing AI math milestones walks through how these numbers are built.
Why the market is moving now
The clearest driver is timing. The 67th International Mathematical Olympiad was held in Shanghai, China, with the competition window running July 10 to 21, 2026, and the two exam papers sat on July 15 and July 16, per the IMO’s official event page. With the exam window now closed, trading has concentrated on verification rather than anticipation, which is typical once an event passes and uncertainty narrows to a single remaining question: can a result be confirmed.
The second driver is momentum, though it has cut both ways. Over the past year, the leading AI labs moved from silver-level to gold-level performance on IMO problems, and traders initially extrapolated that trajectory toward a flawless paper, pushing the price to 97% by July 23. Since then, the price has eased to 77%. It is worth stating plainly: as of this update, no AI perfect score at IMO 2026 has been independently confirmed by the IMO or by any named lab. Huawei, Xiaohongshu, and reports naming OpenAI’s GPT-5.6 Pro have all claimed a perfect 42/42, but a market price of 77%, well short of 100%, reflects real doubt about whether any of those claims will be certified before the July 31 close.
IMO 2026 results: 7 perfect human scores and unverified AI claims
Two new facts now anchor the debate. First, official IMO reporting says just 7 of the 666 human contestants who sat the 2026 exams scored a perfect 42, a hit rate of roughly 1% and a reminder of how rare a flawless paper is even among the world’s strongest young mathematicians. That is in the same range as 2025, when just five of roughly 630 contestants managed it.
Second, several AI perfect-score claims have surfaced, but at the time none carried independent confirmation. Huawei’s Celia model and Xiaohongshu’s dots-note 3.0 each reportedly achieved a perfect 42/42, per TechXplore, claims that first circulated on social media, including a LinkedIn post cited by the Nigerian outlet Vanguard, before the companies themselves put their names to the result. Chinese outlet 36kr separately reported that Claude Fable 5, GPT-5.6 Sol (xhigh), and Kimi K3 also each hit a perfect 42/42 in July 2026, with wildly different efficiency: Claude Fable 5 took 2.5 hours and roughly $51 in compute, GPT-5.6 Sol (xhigh) finished in 3.8 hours for about $20, and Kimi K3 needed 17.4 hours at a cost of $31. Neither the IMO nor an independent grader had certified any of these results at the time, and none of the three labs the Manifold market originally tracked most closely for this question, OpenAI, Anthropic, and Google, was the source of the loudest public claims so far — Huawei and Xiaohongshu were. That mismatch between who traders were pricing and who was actually claiming a perfect score is likely part of why the market cooled rather than converging on 100% before its close.
Odds by AI lab: who traders favor
Beyond the single yes-or-no question, the same Manifold market page priced which specific lab was most likely to land the flawless paper first. As of July 23, 2026, traders put OpenAI at 96%, Anthropic at 94%, and Google at 77% to be the one that achieves a perfect score. That ordering was notable: Anthropic did not publicly headline a gold-medal IMO 2025 result the way Google DeepMind and OpenAI did, yet traders ranked it ahead of Google on this specific question, a sign that the market was pricing recent momentum and rumor as much as the confirmed 2025 track record.
None of these lab-specific odds should be read as confirmation, then or now. They were still probabilities below 100%, meaning traders themselves were not treating any single lab’s result as settled. Notably, the loudest public perfect-score claims to surface since, from Huawei and Xiaohongshu, did not come from any of these three labs at all, which helps explain why the broader “Yes” question has since cooled to 77% rather than rallying toward certainty.
What happened at IMO 2025, the precedent everyone is citing
The 2026 debate rests heavily on what happened one year earlier. In July 2025, Google DeepMind’s advanced Gemini with “Deep Think” scored 35 out of 42 points on the official IMO 2025 problems, solving five of the six questions and clearing the gold-medal threshold. DeepMind said its model worked end to end in natural language within the contest time limit, and it published the result on its official blog.
OpenAI reported the same 35-point score with an experimental general-purpose reasoning model, as covered by Axios and Nature. Both systems reportedly nailed problems one through five and scored zero on problem six. That last detail is central to the perfect-score question: getting from 35 to 42 means solving exactly the kind of problem that stopped both models cold in 2025.
Who and what is involved
The named players include the same labs that reached gold in 2025, plus several new names the market did not initially focus on. Google DeepMind entered an officially graded version of Gemini Deep Think, and OpenAI evaluated an experimental reasoning model on the same problems in 2025. Manifold’s lab-by-lab odds, as of July 23, added Anthropic to the top tier at 94%, alongside OpenAI at 96% and Google at 77%, even though Anthropic did not headline a public gold-medal claim in 2025. In the weeks that followed, the actual public perfect-score claims for 2026 came from elsewhere: Huawei’s Celia, Xiaohongshu’s dots-note 3.0, GPT-5.6 Sol (xhigh), Kimi K3, and Claude Fable 5, none of them certified by the IMO. By August 2026, though, benchmark tracker BenchLM.ai’s own IMO 2026 leaderboard put Anthropic’s Claude Opus 5 in first place with a confirmed 42/42, well ahead of second-place Apodex 1.1, which scored 31 of 42.
Context matters for a human comparison. As CBS News reported on the 2025 event, only about 26 of roughly 630 student contestants, near 4%, outscored the AI systems, and just five students earned a perfect 42. Official IMO 2026 reporting shows a similar rarity: just 7 of 666 human contestants, about 1%, scored a perfect 42 this year. A perfect AI paper in 2026, if ever certified, would put a model in that same tiny top group. For a broader look at how the frontier models stack up, see our comparison of ChatGPT, Claude, Gemini and DeepSeek in 2026.
The skeptics: methodology and verification
Not everyone reads the 2025 results as a clean apples-to-apples win, and that skepticism is part of why a perfect-score claim could be contested even after a lab announces one. IMO president Gregor Dolinar, while confirming DeepMind’s 35-point result, said the organization “cannot validate the methods, including the amount of compute used or whether there was any human involvement, or whether the results can be reproduced.”
Mathematician Terence Tao, a former IMO gold medalist, cautioned on Mathstodon that AI performance depends heavily on the testing setup, with a wide spread in capability depending on the resources a model is given and how results are reported. He said he would not comment on self-reported results whose methodology was not disclosed in advance. Scientific American summarized the broader unease among mathematicians. The takeaway for traders: how a 2026 perfect score is produced and graded may matter as much as the raw number, which is exactly the kind of self-reported, methodology-light claim now surrounding Huawei, Xiaohongshu, and GPT-5.6 Pro — and likely why those claims have coincided with the market falling to 77% rather than rising toward 100%.
How this fits the wider prediction-market picture
The IMO question is one node in a growing cluster of AI-capability markets. A related Polymarket contract on whether an AI wins an IMO gold medal in 2026 has traded at a lower implied probability than the perfect-score market, which is consistent with gold being an easier bar than a flawless paper. Prediction platforms have increasingly become a real-time gauge of where the crowd thinks AI progress is heading, from model release timing to benchmark milestones.
Manifold in particular hosts a long tail of these offbeat, event-driven questions. We have covered similar contracts before, including the Nolan “Odyssey” title-drop market and the Pogacar 2026 Tour de France odds. The pattern is consistent: a specific, checkable outcome, a near-term resolution date, and a price that moves fast as the event approaches.
Timeline: AI at the math olympiad
| Date | Event | Result or figure |
|---|---|---|
| July 2024 | DeepMind AlphaProof and AlphaGeometry on IMO 2024 problems | Silver-medal level (28 of 42) |
| July 2025 | Gemini Deep Think, official grading | Gold, 35 of 42 (5 of 6 solved) |
| July 2025 | OpenAI experimental reasoning model, self-reported | 35 of 42 (5 of 6 solved) |
| July 10-21, 2026 | IMO 2026 held in Shanghai; exams sat July 15-16 | Exams complete; 666 human contestants competed |
| July 13, 2026 | Manifold “perfect score” market | 91% implied “Yes” |
| July 23, 2026 | Manifold “perfect score” market, plus lab breakdown | 97% implied “Yes” overall (peak); 96% OpenAI, 94% Anthropic, 77% Google |
| Late July 2026 | Official IMO 2026 results published | 7 of 666 contestants scored a perfect 42 |
| Late July 2026 | Huawei, Xiaohongshu claim AI perfect scores; GPT-5.6 Pro also named | 42/42 claimed by each; none certified by IMO |
| July 31, 2026 | Manifold market scheduled close | 77% implied “Yes”; resolution due |
Figures for 2024 and 2025 are drawn from the sources listed below; the 2026 rows combine official IMO reporting, market pricing, and unverified AI claims that have not been independently certified.
What to watch next
With the exam window closed, two things will decide this market before its July 31 close. First, disclosure: whether OpenAI, Anthropic, Google, Huawei, Xiaohongshu, or another lab publishes a result with methodology attached, which is exactly what critics asked for after 2025, rather than letting a claim spread through secondhand social posts and company statements alone. Second, verification: a claimed 42 that the IMO or independent graders will not certify could leave the market disputed even after a lab declares victory, which is likely why the price has fallen to 77% rather than climbing toward 100% even with multiple public claims already on the table.
For a tech or finance reader, the signal is less about one contract and more about pace. Each jump on this benchmark, from silver in 2024 to gold in 2025 to a crowd-favored but now-contested perfect score in 2026, compresses the timeline that markets are pricing for advanced reasoning, even as the pullback from 97% to 77% shows traders are demanding real proof before paying up further. That is the same pattern showing up in adjacent bets, such as our coverage of the Claude Sonnet 5 release odds.
FAQ
Has an AI actually scored a perfect 42 at IMO 2026?
At the time of the market’s close, no. The IMO 2026 exams were sat July 15 and 16 in Shanghai. Huawei’s Celia, Xiaohongshu’s dots-note 3.0, GPT-5.6 Sol (xhigh), Kimi K3, and Claude Fable 5 all had publicly claimed a perfect 42/42 by late July 2026, but neither the IMO nor any independent grader had certified those claims. The Manifold price reflected that doubt, peaking at 97% on July 23 before easing to 77% as the July 31 resolution neared. By August 2026, independent tracker BenchLM.ai’s public leaderboard showed Claude Opus 5 posting a confirmed 42/42, with Apodex 1.1 a distant second at 31/42.
How many humans got a perfect score at IMO 2026?
Official IMO reporting says 7 of the 666 human contestants scored a perfect 42, a rate of roughly 1%, similar to the five of about 630 who did so in 2025.
What is the difference between a gold medal and a perfect score?
Gold is a threshold that varies by year and typically requires solving five of six problems; in 2025 it was 35 points. A perfect score is all 42 points, meaning every problem solved, including the two hardest that most human medalists miss.
Which AI models are the likely contenders?
Google DeepMind’s Gemini Deep Think and OpenAI’s reasoning models are the leading names from 2025’s gold-level results, and Manifold’s lab-specific odds as of July 23 put Anthropic in the top tier too, at 94%, ahead of Google’s 77%. But the loudest public perfect-score claims for 2026 have actually come from Huawei and Xiaohongshu, plus reports naming OpenAI’s GPT-5.6 Pro, none of them yet certified.
Why do experts urge caution?
IMO president Gregor Dolinar said the organization cannot validate the compute used, possible human involvement, or reproducibility of AI results. Terence Tao warned that testing methodology can swing reported performance by a wide margin, and that same caution applies to the self-reported claims now circulating from Huawei, Xiaohongshu, and GPT-5.6 Pro.
Is Manifold a real-money market?
Manifold primarily uses play-money. Prices still aggregate crowd belief, but the stakes differ from cash-settled venues, which is worth keeping in mind when reading the odds.
The Bottom Line
Sources
- Manifold Markets — Perfect score achieved by an AI model in the International Math Olympiad (IMO) 2026?
- IMO Official — International Mathematical Olympiad 2026 (Shanghai)
- Google DeepMind — Advanced Gemini with Deep Think achieves gold-medal standard at the IMO
- Axios — In math Olympiad, OpenAI and Google reach gold level
- Nature — DeepMind and OpenAI models solve maths problems at level of top students
- CBS News — Humans triumph over AI at annual math Olympiad, but the machines are catching up
- Scientific American — Mathematicians Question AI Performance at International Math Olympiad
- Terence Tao — Mathstodon post on measuring AI competition performance
- Polymarket — AI wins IMO gold medal in 2026?
Prediction markets carry risk and are not investment or betting advice. Market prices can move sharply and may not reflect actual outcomes. Availability is restricted by jurisdiction: Polymarket is not available to US persons, and Kalshi is a CFTC-regulated US exchange. 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 confidential help.
