Highest OpenAI score on Humanity’s Last Exam in 2026?: how this market works

What you need to know

This market asks how high an OpenAI AI model can score on a benchmark called Humanity's Last Exam (HLE) before the end of 2026. HLE is a very hard test, built from thousands of difficult questions across science, math, law, and other fields, designed specifically to challenge even the most advanced AI. There are three separate yes/no questions here, each with a different score threshold: 50%, 55%, and 60% correct. A 'Yes' on any one means an OpenAI model crossed that line. Each version resolves Yes if any OpenAI model appears on the official HLE leaderboard at agi.safe.ai with an accuracy score at or above the threshold, at any point before December 31, 2026 at 11:59 PM Eastern time. The score that matters is specifically labeled 'HLE Accuracy', not any other metric like calibration error. One important edge case: if that leaderboard goes offline and stays offline permanently, the market resolves No by default, unless official results appear somewhere else. No relevant news was provided for this market. The news item listed appears to be a poem, unrelated to AI benchmarks. The kind of development worth watching here would be OpenAI releasing a new model and submitting it to the HLE leaderboard, or any public announcement of benchmark results from that exam. AI progress on benchmarks like this has surprised people repeatedly in both directions, which is what makes these thresholds genuinely hard to call. The market currently prices the 50% threshold at around 63%, the 55% threshold at 41%, and 60% at 30%, reflecting real doubt at each step up. Key unknowns include what models OpenAI releases this year, whether they submit them to HLE at all, and how quickly AI reasoning improves. A model could leap past one threshold and still fall short of the next.

The odds right now

• 50%+61%
• 55%+40%
• 60%+32%
• 65%+26%
• 70%+8%

Price history

How this resolves

Resolves December 31, 2026

This market will resolve to "Yes" if any model published by OpenAI achieves at least the specified accuracy on Humanity’s Last Exam by December 31, 2026, 11:59 PM ET. Otherwise, this market will resolve to "No". Read the full resolution rules on the live market page.

Related

Same markets. A fraction of the fee.

These apps all route to the same exchange order book. The difference is what each one adds on top of the exchange's own fee.

Published rates, checked July 2026. MetaMask charges a flat 4 percent per prediction trade. Jupiter adds a fee equal to the exchange's own taker fee at fill time, roughly 2 to 4 percent at typical odds. Where a market carries an exchange settlement fee, it applies everywhere, whichever app you use. Paridesk adds nothing on maker orders.

Trade this market on Paridesk: non-custodial, 0.5% fee.