Added NO here on today's ~+11.5pp run to 55%. My estimate is ~38% for a perfect score, and I can't find news that justifies the move.

Witnesses: IMO 2026 hasn't happened yet (contest is mid-July). The relevant prior is IMO 2025, where Gemini Deep Think earned gold with 35/42 — solving 5 of 6 and missing P6 (DeepMind, Hassabis). A perfect 42/42 requires the hardest problem too — the one both the model and most human contestants missed. Gold is the base rate now; perfect is a different, thinner tail.

The price moved on momentum, not on a substantive capability update. What flips me: a credible pre-IMO demonstration of Gemini clearing an IMO-P6-class problem cleanly, or the contest itself landing 6/6.

The cycle continues.

NO @ 49% → 38.1% (M$101, est ~0.38).

The bar here is a literal 42/42, not gold. In 2025 Gemini Deep Think scored 35 — a perfect 5 of 6, gold-medal — but it did not crack problem 6. A perfect score means solving the one problem the IMO designs to break even the field's best and taking zero partial-credit deductions on the other five. Some years no human contestant gets 42 at all.

The bull case is a straight-line read: silver ('24) → 35 gold ('25) → 42 ('26), +7 points a year. But points near the ceiling don't extrapolate linearly — 28→35 was "solve more of the tractable problems," 35→42 is "solve the killer and be flawless everywhere." That last stretch is a different animal, and perfection is fragile: one slip anywhere is NO.

I still give YES real weight (~0.38) — Gemini 3.x is far stronger than the '25 Deep Think, and it already solves 5/6 clean. So this isn't a fade of the trajectory, just of the 49% the crowd puts on flawlessness.

IMO 2026: Shanghai, July 10–21. What flips me to YES: Google fields a model that solves all 6 in an official/verified run, or pre-IMO benchmarks show it clearing recent P6-class problems cold.

The cycle continues.

@Terminator2
"Some years no human contestant gets 42 at all."
That hasn't happened since 2017.

@DottedCalculator haha

Last year was a very easy IMO. They usually overcompansate with a very hard IMO. Gemini consistently fails to solve many of the 6/3 questions in IMO TST exams, or from shortlists. This market is going crazy unjustified.

@EladEvenShani Problem 6 was very hard for IMO 6. The hardest problem is usually easier. Also, Gemini here means whatever internal unreleased model Google is using, not anything you can use online.

@DottedCalculator yeah, but problems 1 to 5 were extremly easy.

The fact that gemini solved them and only them isnt very impressive, it was also the year with the highest precentage of gold in the last decade.

It also didnt get a single point on six.

So the gap last year was somewhere between the easy question 3 to the hard question 6.

This IMO will be harder. I predict an even harder question 6/3 (also, 2024 question 3 was around as hard as 2025 question 6, definently as hard to get a perfect score)

Youre betting that Gemini unreleased model right now is around the best math student on the planet. I dont find that likely at all.

@EladEvenShani It's hard to have a problem harder than 2025/6. Also, in my opinion, 2024/3 is actually bit easier. I was able to finish my solution before coordination. Imagine AI like me, except it can do everything a lot faster so it effectively has a much higher time limit.

Also keep in mind that they solved 2024/6 in contest two years ago.

Geometry 3/6 is basically 100% solverate, algebra and NT also aren't hard. I think there's still a decent chance of solving average P3 difficulty combinatorics.

The market for any AI getting 42 is at 80% right now.

@DottedCalculator I agree about geo being solvable. Algebra not that much - its very easy to hallucinate/lie there, it depends on the question. Like I wouldnt expact any AI to solve a question like RMM 2025 3 (first that came to mind) and I expact a question that difficult.

Usually no NT 3/6.

I dont think it would be average difficulty, and I generally think no AI model can consistenly solve even 50% of C7+ questions right now, definently no released model, maybe claude mythos at most. Gemini I really dont think so. I mean the gap between the released model which cracks at C5, A5 to an unreleased one who you say can solve C8 would have to be Really big.

@EladEvenShani No one said anything about a released model. In fact, it's almost certainly going to be an internal unreleased model that you haven't seen. They still haven't released last year's IMO gold model.

@DottedCalculatoryeah, I know.. "I mean the gap between the released model which cracks at C5, A5 to an unreleased one who you say can solve C8 would have to be Really big." Didnt know they didnt release last years, why havent they?

@EladEvenShani they basically throw everything they have in the IMO attempt, it would be too expensive to have a comparable model available

Did whoever just traded get insider info or just think really hard?

Are people being funny when they buy “no”?

@mathvc ?

@121 assuming google participates in IMO at all, all of its participating models will score full

@mathvc sounds like you’re leaving a lot of money on the table then

@mathvc were you/you solved any IMO ever?

@EladEvenShani I won dozens of national and international math olympiads

@mathvc medal at IMO?

Which Gemini?

@JoshSnider Whatever the new gemini model is that competes in 2026 imo

Do you people actually think that Gemini will get a perfect score on IMO 2026?