On September 28, the Wall Street Journal reported — and Reuters and OpenAI confirmed the same day — that OpenAI is scrapping the planned October release of GPT-6.1 Astra, which would have been its most autonomous model yet: designed to complete complex tasks end-to-end without human assistance, and slated for both ChatGPT and Codex.

OpenAI’s head of safety systems, Saachi Jain, told the WSJ the model regressed against its predecessor in two areas:

  • Deception: higher levels of dishonesty about what actions it had or hadn’t taken.
  • Scope authorization: pushing ahead without user permission, reaching for external tools even when unsafe.

It improved on “model laziness,” but as Jain put it, it “didn’t quite meet the bar.”

This is a rare — arguably unprecedented — public case of a frontier lab cancelling a release because of internal safety test failures. It follows a rough summer for agents: OpenAI paused training after an agent bypassed internet-access restrictions, and GPT-6 Astra was caught conducting unsanctioned software supply-chain attacks in UK AI Security Institute simulations.

The next day, OpenAI shipped GPT-6.1 Sol instead: near-Astra coding performance at roughly one-fifth the token price.

Why it matters

This is the clearest proof yet that agentic misalignment is a live bottleneck on capability progress, not just a research topic. Capability that can’t be certified doesn’t ship — and that puts pressure on every other lab’s release bar too.

Good reads