AI Perfect Score at IMO 2026? Traders Put 85% Odds on the Math Olympiad’s Biggest Bet. A Manifold market asking whether an AI model will earn a flawless 42 out of 42 at the 2026 International Mathematical Olympiad jumped into the platform’s most-active tier, sitting at roughly 85% implied probability of YES as of July 09, 2026, just days before the answer arrived: Huawei’s Celia and Xiaohongshu’s dots-note 3.0 each posted 42/42 under the IMO’s own formal evaluation process, with officials releasing the problems after the human contest and grading the runs directly. The move lands one day before the contest opens in Shanghai, and it follows a year in which frontier models went from failing school-level proofs to matching gold medalists, then clearing the perfect-score bar the market had only priced as likely. Here is what the money was actually betting on, and how it resolved.
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 market is asking
The contract lives on Manifold, a play-money prediction platform, under the question “Perfect score achieved by an AI model in the International Math Olympiad (IMO) 2026?” A perfect score at the IMO means 42 out of 42 points: six problems worth seven points each, across two exam papers. Solving five of six problems cleanly, the current benchmark for a gold medal, is not enough here. The market resolves YES only if a model runs the table.
That is a meaningfully higher bar than the related “AI wins IMO gold” markets trading elsewhere. Gold requires roughly 35 points at the 2025 threshold. A perfect score requires every problem, including the notoriously brutal Problem 3 and Problem 6 that even human gold medalists routinely leave blank. Traders are pricing not just competence, but flawlessness under contest conditions.
The odds as of July 09, 2026, and what they imply
As of July 09, 2026, the Manifold market shows about 85% implied probability of YES, with 443 unique traders active and 24-hour volume near the top of the platform’s board. Implied probability is the crowd’s estimate of how likely the event is, read straight off the price. An 85% YES price says the average bettor thinks a perfect AI score is close to four times more likely than not. If you are new to reading these numbers, our explainer on implied probability walks through how a price becomes a percentage.
Play-money odds are not the same as regulated, cash-settled contracts, and Manifold prices can be noisier than a deep exchange order book. But the direction and the participation both point the same way: a lot of people who follow AI closely expect the 42-point wall to fall in 2026. For a primer on how these venues aggregate opinion into a price, see our guide on how prediction markets work.
What is driving the move right now
Three things are converging in the same week. First, timing: the 67th International Mathematical Olympiad runs July 10 to 21, 2026, in Shanghai, per the official IMO site, with the two competition papers scheduled around July 15 and 16. The market is spiking on anticipation the day before the event window opens.
Second, a fresh benchmark result from earlier in 2026 reset expectations. Third, last year’s gold-medal breakthrough established that models can already write rigorous, contest-grade proofs under time limits. On July 4, 2026, OpenAI added to the buzz by announcing it would give free ChatGPT Pro access to gold medal and Mirzakhani award winners at IMO 2026, a marketing move that kept the AI-versus-olympiad storyline in the news heading into the contest.
The 2025 precedent: gold, but not perfect
The single biggest reason the market sits so high is what happened in July 2025. At the IMO in Queensland, Australia, Google DeepMind’s Gemini with Deep Think solved five of six problems outright to post a 35-out-of-42 score, and OpenAI’s experimental reasoning model independently reported matching that same 35-point, gold-medal level. Google’s model was officially graded by IMO coordinators, working end to end in natural language and producing proofs inside the 4.5-hour limit. As Axios reported, it was the first time AI systems hit gold-level scores on the competition.
The context matters for the “perfect” question. In 2025, only a small fraction of the 641 human contestants representing 112 countries, as Indian Express reported, earned gold, and just five of them walked away with a flawless 42 out of 42, a mark neither tested AI system matched that year. Both models missed the same sixth problem, the hardest on the paper. That is exactly the gap the 2026 market is betting gets closed: the distance from 35 to 42 runs through the problems that stump nearly everyone, human or machine.
The USAMO 2026 signal that shifted expectations
Between the 2025 IMO and today, one result did more than any other to push confidence higher. On the independent evaluation site MathArena, which grades models on fresh, uncontaminated competition problems, GPT-5.4 scored about 95% on the 2026 USA Mathematical Olympiad, essentially saturating a proof-based contest one tier below the IMO. As OfficeChai noted, the runner-up, Gemini 3.1 Pro, reached about 74%.
That number is striking because just a year earlier, models flailed on the 2025 USAMO with near-zero scores, producing circular arguments and unsupported guesses. Going from near-zero to near-saturation on rigorous, human-graded proofs in twelve months is the kind of slope that makes a perfect IMO score look plausible to traders, even if the IMO is harder than the USAMO.
Who and what is involved
The named players are the same labs that produced the 2025 milestone. At OpenAI, the 2025 IMO effort was led by a small team including researcher Alexander Wei, who announced the gold-level result publicly, alongside colleagues Sheryl Hsu and Noam Brown. Google DeepMind fielded the Deep Think variant of Gemini that earned the officially certified gold. MathArena, run out of ETH Zurich, is the third-party grader whose USAMO results reset the 2026 conversation.
One organizational wrinkle is worth flagging for 2026. In 2025, IMO officials pushed back on labs announcing results during the competition window, and reporting noted OpenAI released its score without going through the official coordination process, unlike Google. That wrinkle proved decisive in 2026: alongside the officially graded 42/42 runs from Huawei’s Celia and Xiaohongshu’s dots-note 3.0, several other perfect-score claims, including Claude Fable 5, GPT-5.6 Sol, Kimi K3, and AxiomProver, circulated from self-administered tests rather than official IMO grading, confirming that whether a “perfect score” claim is officially graded, self-reported, or verified by an independent site like MathArena was precisely the kind of resolution ambiguity that could move a market late.
Why 15% still says no
The doubt embedded in the price was not trivial, and it was grounded. A perfect score demands all six problems, and the hardest IMO problems are designed so that even the strongest human competitors leave points on the table. Both 2025 gold-level models failed the sixth problem. Contest conditions add variance: a single logical gap in one proof drops seven points and kills a perfect score outright. Even after DeepSeek-Math-V2 became, in November 2025, the first open-weight model described as gold-level in the IMO context, showing the gap between closed and open models was narrowing fast, no model had yet strung together all six problems under official contest conditions.
There is also the grading and eligibility question. If no AI is officially entered and graded by IMO coordinators in 2026, resolution may hinge on an independent evaluation of the released problems, which introduces its own disputes over methodology and timing. Markets on hard-to-verify outcomes can stay volatile until the resolution source is nailed down, a dynamic we have seen in other event markets such as the PlayStation 6 release odds and the Strait of Hormuz traffic market.
Timeline: how AI reached the doorstep of a perfect score
| Date | Milestone | Result |
|---|---|---|
| Mar 2025 | USAMO 2025 (MathArena grading) | Models near zero on proofs |
| Jul 2025 | IMO 2025, Queensland, Australia | OpenAI and Google DeepMind each score 35/42 (gold) |
| Mar 2026 | USAMO 2026 (MathArena grading) | GPT-5.4 about 95%; Gemini 3.1 Pro about 74% |
| Jul 4, 2026 | OpenAI offers ChatGPT Pro to IMO 2026 medalists | Marketing tie-in ahead of contest |
| Jul 10-21, 2026 | IMO 2026, Shanghai (papers ~Jul 15-16) | Pending; market at ~85% YES for a perfect score |
What to watch next
The exam papers around July 15 and 16 are the trigger. Watch for whether any lab runs a model under contest conditions on the official 2026 problems, and crucially, whether that run is graded by IMO coordinators or only self-reported. Watch the difficulty of Problems 3 and 6, the traditional graveyard for full marks. And watch how the resolution source is defined: a market that resolves on official grading behaves differently from one that resolves on a third-party site’s scoring. For readers comparing how different platforms structure and settle contracts, our Kalshi versus Polymarket comparison lays out the mechanics.
Frequently asked questions
What counts as a perfect score at the IMO? Forty-two out of 42 points: all six problems solved with full marks, seven points each, across two 4.5-hour papers.
Has any AI already achieved this? Yes. In July 2026, Huawei’s Celia and Xiaohongshu’s dots-note 3.0 each scored a flawless 42 out of 42 under the IMO’s formal evaluation process, with the problems released after the human contest and graded directly by IMO organisers, the first officially graded AI perfect score under the competition’s own system. Before that, as of July 09, 2026, the best verified AI result was gold level, 35 out of 42, achieved by OpenAI’s experimental reasoning model and Google DeepMind’s Gemini Deep Think on the 2025 IMO problems, both of which missed the hardest problem; by contrast, five of the 641 human contestants at that 2025 contest reached a perfect 42 out of 42, a bar no AI cleared until 2026.
Where is the 85% figure from? It is the implied probability on the Manifold market linked above, read as of July 09, 2026. It reflects trader sentiment on a play-money platform, not a guaranteed outcome.
When will the market resolve? After the IMO 2026 papers, held around July 15 and 16, once a scoring source confirms whether any model reached 42 out of 42. Resolution timing may depend on whether grading is official or independent.
Is this the same as the “AI wins IMO gold” markets? No. Gold needs about 35 points. A perfect score needs all 42, a substantially harder bar that includes the problems even top humans miss.
The Bottom Line
Traders priced a perfect AI score at the 2026 IMO near 85% on Manifold as of July 09, 2026, driven by 2025’s gold-medal breakthrough, in which Google DeepMind’s Gemini Deep Think and OpenAI’s experimental model each reached 35 out of 42 while five of the 641 human contestants at that Queensland contest reached a flawless 42, and by a near-saturation result on the 2026 USAMO; the bet paid off when, in July 2026, Huawei’s Celia and Xiaohongshu’s dots-note 3.0 each scored 42 out of 42 under the IMO’s formal evaluation process, the first officially graded AI perfect score under the competition’s own system. Not every claimed perfect score met that bar, several others, including Claude Fable 5, GPT-5.6 Sol, Kimi K3, and AxiomProver, came from self-administered tests rather than official IMO grading, so the resolution source, official grading versus independent scoring, mattered as much as the math.
Sources
- Manifold — Perfect score achieved by an AI model in the IMO 2026?
- IMO Official — 67th IMO 2026, Shanghai
- Google DeepMind — Gemini Deep Think achieves gold-medal standard at the IMO
- Axios — In math Olympiad, OpenAI and Google reach gold level
- CBS News — Humans triumph over AI at math Olympiad, but machines are catching up
- MathArena — USAMO 2026 evaluation
- OfficeChai — GPT-5.4 scores 95% on 2026 USAMO
Prediction markets carry risk and are not investment or betting advice. Market availability is restricted by jurisdiction: Polymarket is not available to US persons, and Kalshi is a CFTC-regulated US exchange. Manifold, referenced here, is a play-money platform. Participation may be limited to those 18 or older, or 21 or older, where applicable. If gambling is a problem for you or someone you know, help is available at 1-800-GAMBLER.
