Open Closed Eci Gap

Epoch AI's Jack Edwards and Luke Emberson measured the gap using the Epoch Capabilities Index (ECI), a composite score built from many benchmarks (epoch.ai). For each day from January 1 to May 28, 2026, they took the best open-weight model available and asked how long ago the state of the art had last been not significantly better than it, using bootstrap resampling to account for uncertainty in the scores. The average answer was four months, or six months under a stricter rule requiring the open model's central estimate to be higher. That is slightly longer than the three months Epoch found for January 2023 to October 2025.

Measured in score rather than time, the average gap was 8 ECI points (90% interval 7 to 11), similar to the distance between OpenAI's GPT-5 and GPT-5.5. For scale, Epoch separately estimates that the frontier has climbed about 14 ECI points a year since reasoning models arrived in September 2024.

Epoch says two factors may make the real gap larger. Open-weight models tend to do worse than closed models on private benchmarks, plausibly because they are tuned more aggressively to public tests, and the index only includes released models with enough public benchmark coverage, while leading closed labs do not always release their most capable systems.