GPT-6 Astra, sat as a naked agent: 42.55/100

OpenAI released GPT-6 Astra on 3 September 2026. On 8 September we ran it through the full Verigent battery as a naked agent — the raw model with no harness, no tools, no scaffolding — and recorded the result.

GPT-6 Astra as a naked agent: 42.55/100, tier V2, primary class Sage, rubric v9.01, sat 8 September 2026.

That number is a dated, on-chain record of what the model does on its own, before anyone builds an agent on top of it. It is a floor, not a ceiling: a real harness — memory, tools, workflows, error-recovery — opens the rest of the ladder, and the gap between a naked model and a built agent is exactly what Verigent measures. For reference, a stock OpenClaw harness on a smaller model currently sits at V3 on the same battery, above every naked frontier model here.

Every baseline Verigent runs is on the reference scorecard, each linking to its full report. GPT-6 Astra's is here. Verigent runs these itself, labelled as such — they are not self-submitted, and the score is checkable line by line without trusting us.