Reasoning models more than doubled the pace of AI progress, Epoch finds

Epoch AI put a number on something people have felt since late 2024: reasoning models sped up the pace of AI progress, and by a lot. Using its Epoch Capabilities Index, which folds many separate benchmarks into one general-capability scale so models from different eras can be compared, the frontier has advanced about 14 ECI points per year since reasoning models appeared in September 2024. Before that, with non-reasoning models, the rate was roughly 6 points per year.

The gap is a 233% acceleration, or about one point every 26 days. Epoch tracks the line across specific checkpoints: o1-mini at 136.64 in September 2024, o1 at 142.67 that December, up to Claude Fable 5 at 162.48 in June 2026 and GPT-5.6 Sol at 161.06 in July. Across 13 frontier reasoning models released through August 2026, the climb stays close to linear, which is the striking part. A steady line, rather than a flattening curve, is what argues against near-term diminishing returns.

The caveat is built into the method. ECI blends benchmarks, and benchmarks saturate, so a single index can hide where specific capabilities stall. Still, the trend break at September 2024 is clear in the data.

Why it matters

If you plan roadmaps or budgets around model capability, a steady 14 points a year is a concrete rate to plan against instead of vibes. It also raises the bet that the next year keeps climbing at that pace, which changes how long you should expect any given model to stay competitive.