Epoch AI: open models lag closed frontier by four months
The gap, measured on Epoch's Capabilities Index, has held at roughly eight index points since January 2026 — comparable to the distance between GPT-5 and GPT-5.5.
- Benchmarks & progress
- Open weights & ecosystem
- Minor
Epoch AI reported that the most capable open-weight models have trailed the closed-model frontier by an average of about four months since the start of 2026, measured on the organisation’s own Epoch Capabilities Index (ECI). It put the typical gap at roughly eight ECI points — a distance it compared to the difference between OpenAI’s GPT-5 and GPT-5.5 — and said this spread had stayed fairly stable across the first five months of the year rather than widening or closing.
At the time of the analysis, the leading open-weight models by ECI score were Kimi K2.6, GLM-5.1 and Qwen 3.6 Plus, against a closed-model frontier led by GPT-5.5 Pro, Claude Opus 4.7 and Gemini 3.5 Flash. Because ECI aggregates performance across many benchmarks into a single score, the four-month figure is a summary measure rather than a claim about any specific capability; it says nothing about gaps that might be larger or smaller in particular domains such as coding or long-horizon agentic tasks.
The finding bears on a recurring argument about whether open-weight releases are converging with or falling behind proprietary frontier labs: a roughly stable multi-month gap, sustained through a period of frequent releases on both sides, suggests neither has been closing on the other in early-to-mid 2026.