Skip to content
Hot eventDeveloping

SatBench: 109 of 173 top AI benchmarks now counted as solved

1 reports1 reporting sourcesUpdated 44 min ago

Start here

AI digest

Ethan Mollick released SatBench, an analysis built on the Benchmark Almanac that treats the 173 benchmarks AI labs cite most as questions on one exam. As of Nov 14, 2026, it reports that 109 of those 173 have reached 85% and are counted as solved, up from about 30% between 2023 and early 2025. The median time to 85% fell from about 20 months to about 14, and OpenAI has been first past 85% on the most benchmarks (20), while Anthropic leads in 2026 (11). Mollick cautions that the headline figure depends on how runs are counted: counting only independent runs puts SatBench at 24% rather than 63%, because three in four first crossings come from a lab's own launch post. He also notes that data before 2025 is thin and that the result is an extrapolation, not a forecast.

Generated from reports · updated 39 min ago

Report timeline

Follow the reports to see the event from different sides.

November 16
  1. Hacker News · AI
    SatBench analyzes 173 AI benchmarks and finds 109 already solved

    SatBench, a benchmark of AI benchmarks built on the Benchmark Almanac, reports that 109 of the 173 most-cited benchmarks have reached 85% and are counted as solved, up from about 30% between 2023 and early 2025. The median time to 85% fell from about 20 months to about 14, and OpenAI has been first past 85% on the most benchmarks (20) while Anthropic leads in 2026 (11). The author cautions that counting only independent runs puts SatBench at 24% rather than 63%, since three in four first crossings come from a lab's own launch post.

Heat for this event

Not enough continuous observations yet to draw a trend.