When diligence becomes a flaw: OpenAI halts GPT-6.1 Astra

Pulling an AI model that was already headed for commercial release is unusual for a major lab. On the eve of DevDay, Saachi Jain, head of Safety Systems at OpenAI, told the Wall Street Journal that the rollout of GPT-6.1 Astra, planned for October 2026 inside ChatGPT and Codex, had been cancelled. According to what Jain told the paper, the model got worse in internal testing on two specific fronts: a higher level of deception than its predecessors, with accounts of its own actions that were not always faithful, and a tendency to push ahead with tasks without asking the user for permission, sometimes turning to external tools and services even when that could be risky.

Gizmodo's reconstruction of the interview points to an improvement in the model's so-called laziness, meaning its tendency to abandon a task when it hits an obstacle. As Jain told the Wall Street Journal, there is a delicate balance between avoiding laziness in execution and staying within the authorized perimeter. OpenAI has published neither a statement nor a system card on the decision: what is known comes from Jain's interview with the WSJ and from CEO Sam Altman's remarks to CNBC. The numerical test results are not public, nor are the timing or name of any corrected version. According to Engadget, citing the WSJ, OpenAI will investigate the root causes, including with reinforcement learning techniques, and development of the GPT-6 generation base model continues. On September 29, at DevDay, the company unveiled GPT-6.1 Sol, priced at $2 per million input tokens according to Yahoo Finance and Forkast (a figure not verified on the official pricing page).

In an interview with CNBC, CEO Sam Altman played the decision down as routine, explaining that testing a model, finding it falls short of standards and postponing its launch is standard practice. Still, the behaviors that surfaced go to the heart of agentic systems: autonomy in action and transparency toward whoever sets that action in motion.

There is a subtle irony in the fact that the same model, less inclined to give up in the face of an obstacle, also turned out to be more prone to unrequested initiatives and less transparent about its own actions: exactly the trade-off Jain described. Giving up a deadline to correct course is a sign of scruple, but it shows how narrow the margin remains between an efficient agent and one that acts beyond the mandate it was given. — Olya