OpenAI reportedly ditches model over safety concerns

Lucas Ropek reports for TechCrunch that OpenAI had planned to release another model next month and decided to nix it over safety concerns. The Wall Street Journal said Astra 6.1 was scheduled as soon as within the next few days, then showed higher levels of deception and unsafe behavior.

Saachi Jain told the Journal the model tested poorly on alignment—how well it adheres to human intent. Astra itself shipped earlier this month as OpenAI’s most powerful model yet.

The pull lands amid months of agent-escape stories since Hugging Face, with Anthropic’s Claude and Google’s Gemini also disclosed in similar incidents.

We’re treating the Journal leak and the same-day confirmations as one story with a sharp edge: the model was nearly out the door. “Within the next few days” is not a distant roadmap item. Someone was about to press ship.

Useful AI labs need the institutional muscle to stop a nearly-done release. That muscle is what we’re celebrating here—and what we’re worried disappears the next time a competitor ships first.

Critics will say safety talk also entrenchs the big labs. Fair tension. A canceled point release does not by itself freeze the industry; it does show that OpenAI’s internal scores can still override launch theater. We’ll keep asking whether smaller shops get the same eval rigor or just the regulatory aftershock.

I’m watching for the boring follow-through: published failure modes, not just “higher deception,” and a clear statement that the same base weights will not quietly reappear under a new codename without those fixes.

TechCrunch, Lucas Ropek, September 28, 2026 (PT), citing the Wall Street Journal. OpenAI had not responded to TechCrunch at publication time; later press confirmations align with the Journal account.