Ethan Mollick's count: on 63% of 173 AI benchmarks the best reported score reaches 85%, against 24% when only third-party scores count

Wharton professor Ethan Mollick published SatBench on November 14, a tally of the 173 benchmarks AI labs have cited in their release posts since 2023 (SatBench, 2026-11-14). For three-quarters of the benchmarks that crossed 85%, the first crossing came in a lab's own release post. Counting only scores run by third parties, 24% of all 173 benchmarks reach that mark.
SatBench counts a benchmark as saturated once the best reported AI score reaches 85%, and 109 of the 173, or 63%, meet that bar. Mollick gives the same figure on X (Ethan Mollick, 2026-11-14).
Counting only scores run independently by third parties, 24% of the benchmarks meet the bar. SatBench's documentation says that for three-quarters of the benchmarks that crossed 85%, the first crossing came in a lab's own release post. The documentation also says the self-reported scores mix different harnesses, tool use and best-of-several attempts. A harness is the code wrapped around a model that feeds it the questions and connects its tools.
New benchmarks are being pushed past 85% faster. The share reaching the bar within a year of release was 17% for benchmarks released in 2023, 30% for 2024 and 43% for 2025. Some of the 2025 benchmarks are not yet a year old, so that cohort's share isn't final.
Not proven yet: Mollick chose both the 85% bar and the rule for which benchmarks count How we checked →
Next to watch: whether anyone recalculates from the same public data with a different bar or with independent scores only, and what share of the benchmarks cited in labs' next round of model release posts are new in 2026.
Source and checks
This item ran in the brief of Nov 15, 2026, alongside 9 other items. How we checked it, the quote checks and what was still missing that day are on that issue's How we checked page.
Cite this item
SecondSource (2026). "Ethan Mollick's count: on 63% of 173 AI benchmarks the best reported score reaches 85%, against 24% when only third-party scores count." https://secondsource.io/en/news/2026-11-15-2/. Accessed access date.