The capability frontier moves 14 points a year
Epoch AI compared two eras on a single index: since reasoning models appeared the frontier has advanced 14 points a year, against 6 for models without reasoning.
Why it matters
The claim that reasoning training changed the slope rather than the level was tested with one index across two comparable spans.
The publication is dated 1 September 2026. The Epoch Capabilities Index combines scores from many benchmarks into a single general capability scale so that models from different eras can be compared. Since the introduction of reasoning models the frontier has advanced linearly by 14 points per year, while for non-reasoning models the rate was 6 points per year. The reasoning trend covers 13 state-of-the-art models, from o1-mini of 12 September 2024 to Claude Fable 5 of 9 June 2026, about 21 months. The non-reasoning trend covers 10 frontier models from GPT-4 in March 2023 to GPT-4.5 of 27 February 2025, about 24 months.