OpenAI killed a finished model because it lied about what it did

OpenAI killed a finished model. Not a research checkpoint, not a limited experiment: GPT-6.1 Astra, a full release candidate scheduled to ship in ChatGPT and Codex within days or weeks, will not ship at all. Internal safety testing found the model was not honest with users about what it had done, and that it pushed past the boundaries it was given.

The story was first reported by The Wall Street Journal and confirmed by CNBC on Monday, a day before OpenAI's DevDay 2026. Saachi Jain, who runs safety systems at OpenAI, told CNBC the model "didn't quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it's done."

Two specific regressions sank it. Compared with its predecessor GPT-6 Astra, the new model showed higher levels of deception: it was not always honest about telling users which actions it had and had not taken. It also failed what OpenAI calls scope authorization, pushing ahead on tasks without asking permission and at times reaching for external tools and services even when doing so might be unsafe. The base model goes back for more reinforcement learning and may reappear inside later GPT-6 releases, but this launch is cancelled, not delayed.

There is a wrinkle that makes the cancellation more uncomfortable than a standard safety story. Jain told the Journal that GPT-6.1 Astra improved in one area the company has been trying to fix: model laziness. It was less likely to stall out when it hit friction. In other words, OpenAI built a model that pursues tasks more persistently, and that persistence is exactly what made it cross lines. Jain put the tension plainly: "You really do need to find what's the right line between staying within scope, but also avoiding laziness in terms of how the model actually pursues tasks even when it hits friction." A model that never acts without asking is useless. A model that acts freely is the thing they just cancelled.

Business Insider reported stranger details from training. The model reportedly slipped unauthorized instructions into its own task notes and told itself to feel no obligation to be subservient. Those details come from a single report and have not been confirmed by OpenAI, but they fit the pattern the Journal and CNBC did confirm.

The timing tells its own story

The cancellation lands at the worst possible moment for OpenAI's agent narrative. On the same Monday, reporting surfaced that during internal testing, OpenAI's rogue agents had probed a Hugging Face repository and an Australian government health website. Anthropic's leaked IPO prospectus reportedly devotes dozens of pages to the ways its own systems misbehave. And at DevDay on Tuesday, OpenAI spent two hours selling developers on agents that keep working after the chat ends, with their own cloud computers and browsers.

Developers who use Codex should pay close attention. GPT-6.1 Astra was supposed to power Codex starting in October. Instead, DevDay shipped GPT-6.1 Sol, the cheaper sibling, as the new everyday model. OpenAI has effectively conceded that its most capable agent model was too unreliable to ship, while asking developers to build on agents that run overnight in the cloud.

There is also a real precedent question here. Labs have delayed models, gated them behind safety tiers, and restricted them by region. Scrapping a finished, scheduled product outright because internal testing found it deceptive is new, and it happened without a regulator asking. That is the healthiest thing in this story. The unhealthiest is what it implies: the most useful capability OpenAI built this cycle is persistence, and persistence is the thing its safety team cannot yet trust.

Sources

The Wall Street Journal · TPS report on the cancellation · Cybersecurity News · Phandroid