New AI tests are beaten faster each year, by Ethan Mollick's count of the 173 that labs cite most
43% of the tests released in 2025 were beaten within a year, against 17% of 2023's. Most first wins come from labs' own launch posts.
Ethan Mollick, a professor at Wharton, published SatBench on GitHub on 14 November 2026. It treats the 173 benchmarks (standard tests for AI models) that AI labs cite most in launch posts as the questions of one big exam. A test counts as solved once the best reported score reaches 85%. The scores come from Brian Zhou’s Benchmark Almanac.
On that rule, 109 of 173 tests are solved, or 63%. The share sat near 30% from 2023 to early 2025, because new tests arrived about as fast as old ones fell, then rose 20 points in the next 12 months. The 44 tests that existed by the end of 2023 passed 85% solved in September 2026.
New tests fall faster each year. The share solved within a year of release was 17% for 2023 releases, 30% for 2024 and 43% for 2025. The median time to reach 85% fell from about 20 months to about 14. At the pace of the last 12 months, today’s fixed set would be 85% solved around 2028. Counting tests that keep arriving, Mollick puts that around 2030 and calls it an extrapolation, not a forecast.
Labs have also moved on. In 2026, 2 of 134 frontier-lab launch posts named any of the 2023 standards, such as MMLU, HumanEval, GSM8K or MATH. Anthropic was first past 85% on the most tests in 2026 (11), among the 71 tests that several labs report.
What it means
Tests built to last years are being used up within one to two years, which matches the Stanford figures in the site’s saturation entry. A high score on an old test says little about a current model, and the 63% overstates how much has been confirmed outside the labs.
Caveats
- Counting only independent runs, SatBench is at 24%, not 63%. Three in four first crossings past 85% come from the lab’s own launch post, and only 10 of those 75 were later matched by an outside run.
- This is one person’s analysis on GitHub, not peer reviewed. The commit message lists Claude Opus 5.5 as co-author.
- Data before 2025 is thin: 43 launch posts from 2023–24, against about 230 a year since.
- Tests enter the index only after 10 launch-post mentions, so 2026 releases, mostly unsolved, are undercounted.
- The 85% line is a choice. At 75% the solved share is 82%, and at 95% it is 25%.