Our highest figure: Alnair 1 · Maximum effort, on the AIME 2025 and 2026 problems. Every figure here comes from our internal measurements in the August 2026 campaign, and each carries the name, version and effort the app shows, its confidence interval and, beside it, without a name, the best score published to date.
Three public tests, three figures.
99.0 %
Alnair 1 · Maximum effort
Competition mathematics: 60 problems, 30 attempts each. 95% confidence interval: 97.0 to 100.0.
Flagged, as it is above 98. Counter test by year of publication: 100.0 on the 2025 problems, 98.0 on the 2026 ones.
The best score published to date is higher. Below the median of the five best. Both reference points fall within this figure's confidence interval.
89.6 %
Alnair 1 · Maximum effort
Competitive programming: 182 problems published between January 2025 and April 2025, 5 attempts each, scored by running the code. 95% confidence interval: 85.3 to 93.5.
These problems were published before the campaign: memorization cannot be ruled out.
The best score published to date is higher. Below the median of the five best. Both reference points fall within this figure's confidence interval.
93.9 %
Alnair 1 · Standard effort
Doctoral level science: 198 questions, one answer each. 95% confidence interval: 90.4 to 97.0.
The best score published to date is higher. Below the median of the five best. Both reference points fall within this figure's confidence interval.
Both reference points are anonymous: for each test, the best score published to date and the median of the five best published scores, one score per system, collected on October 8, 2026 from results their own authors published. Their conditions and problem sets are not always ours: these reference points place a figure, they do not rank it.
The figures on this page come from our internal measurements, the August 2026 campaign: more than 39,000 scored measurements on three public tests.
Each figure states how many questions it covers and how many attempts each one received.
Every question counts: a refusal to answer, or a missing answer, is a failure, never a question removed from the count.
Every figure carries a 95% bootstrap confidence interval, drawn question by question.
A score above 98 is presumed to be recall until a counter test says otherwise: the set is split by year of publication, and the flag stays shown beside the figure.