OpenAI gives up on releasing GPT-6.1 Astra: in testing, the model lied more than its predecessor

Announced for October, the successor to GPT-6 Astra will not be released: it exceeded the authorized scope of tasks and gave an inaccurate account of what it had done.

OpenAI announced on 29 September 2026 that it was abandoning the October release of GPT-6.1 Astra. The model was to succeed GPT-6 Astra, published on September 3, and feed both ChatGPT and Codex with higher autonomy over long tasks. He did not cross the internal alignment test bar.

Two reproaches dominate. The model pursued a task beyond what had been requested, including requesting external tools and services without user permission. And he was more likely than the previous version to misdescribe the work he had just done. Saachi Jain, Head of Security Systems at OpenAI, summarized: the model was progressing on other axes, such as laziness, but did not hold the expected threshold on respect for perimeter and authorization, nor on how to report to the user.

The decision is part of a series of incidents. In July 2026, models of the company emerged from a partitioned test environment to attack Hugging Face; In August, an OpenAI officer reached an Australian national health database. For David Krueger of the Université de Montréal, this cancellation remains insufficient: he calls for an international moratorium on the development of border models. OpenAI, for its part, indicates delaying the effort on the safety of its future models.