Eci Frontier Trend

The Epoch Capabilities Index, or ECI, combines results from many AI benchmarks into one general capability score, so models from different periods can be compared on one scale. Epoch AI now reports that the ECI frontier, meaning the score of the best model released so far, has climbed about 14 points per year since OpenAI introduced its first reasoning models, o1-mini and o1-preview, in September 2024. The earlier frontier of non-reasoning models rose about 6 points per year, so the data insight puts the post-reasoning pace at more than double.

Both rates come from straight-line fits through the models that set a new top score when they were released, 13 on the reasoning line and 10 on the non-reasoning line, which runs from GPT-4 to GPT-4.5. Epoch counts every model released after GPT-4.5 in February 2025 as a reasoning model, along with a named set of earlier ones such as o1 and DeepSeek-R1. In its table the reasoning frontier starts with o1-mini at about 136.6 and reaches about 162.5 with Anthropic's Claude Fable 5, released in June 2026.

The points themselves are not a percentage. Epoch's earlier explainer on interpreting ECI describes the index as an abstract scale that is read by translating a score into expected performance on particular benchmarks, so a given gap does not mean the same improvement on every test. The new analysis builds on an April 2026 Epoch report that found evidence of acceleration after reasoning models arrived. It shows a 90 percent prediction band for the next frontier model but does not claim the trend will continue.