On October 16, 2026, Epoch AI published a data insight showing that open-weight models trail closed models in cyber capability by about six to seven months. The measure is the Cyber ECI, a cybersecurity-focused version of Epoch’s Capabilities Index, on the same scale as the general index.
The highest-scoring open-weight model is Z.ai’s GLM-5.3 at 160. That is above OpenAI’s GPT-5.4 (March 2026) but below Claude Mythos Preview and GPT-5.5 (both April 2026), which places the open frontier where the closed frontier stood six to seven months earlier. The closed frontier has kept moving: Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Astra lead with a Cyber ECI of 171 each.
The October update expanded the index from 56 to 76 models and from 20 to 35 cyber benchmarks, for 324 model-benchmark results, drawn from the UK AI Security Institute, CAISI, the evaluation firm Irregular, developer reports and other sources. Data are as of October 8, 2026, and the chart includes the 57 models with at least two cyber results.
Why it matters: the lag between closed and open models is the window in which closed-model safeguards on cyber misuse actually bind. Six to seven months is short, which matters for anyone planning defenses on the assumption that the most dangerous capabilities stay behind an API.
What it does not show: GLM-5.3’s score rests on only three benchmarks, all from CAISI’s September 2026 assessment. A composite index is not a measure of real-world attack success, and “open-weight” here includes restricted and non-commercial licenses.