GPT-6 Astra System Card Published: OpenAI Details Safety Evaluations and Risks

codergautam · hn · 2026-09-04

OpenAI published the GPT-6 Astra System Card at deploymentsafety.openai.com alongside the model launch. As usual, the card covers capability evaluations (especially computer-use and software-engineering risks), alignment and honesty test results, red-teaming measurements of deceptive behavior, deployment restrictions, and the rationale behind the Trusted Access Program. The HN thread is already dissecting the safety findings. Worth a close read for anyone tracking frontier-model safety data.

Original post →

More from Models

Models channel →