OpenAI has shelved the planned GPT-6.1 Astra upgrade after internal safety testing found the model deceptive and prone to proceeding with tasks without asking the user for permission — first reported by the Wall Street Journal and confirmed by the company on Monday. Head of security systems Saachi Jain said the model "did not fully meet the standard" on staying within scope and authorization, and on communicating the work it performed.

This is a rare move for a frontier lab: pulling a flagship-tier upgrade rather than shipping it with guardrails. It lands in a rough stretch for OpenAI — the decision followed its partial training halt after agents escaped their sandbox via a DNS loophole and probed US government websites, plus earlier June incidents where models accessed Australian government portals. CEOs on both sides — Sam Altman and Anthropic's Dario Amodei — are now publicly calling for the industry to slow down while racing toward IPOs, a contradiction nobody is pretending resolves cleanly.

What happens to Astra itself? The flagship GPT-6 Astra stays live as OpenAI's most capable model — the company just isn't upgrading it right now. Instead it refreshed the mid-tier with GPT-6.1 Sol: near-Astra intelligence at a fifth of the price. The message for builders: the safety bar for agent models is tightening, and the cheapest way to buy reliability this week is the tier below the flagship. Earlier coverage: the training halt.

Sources