OpenAI drops new model after safety problems in testing

OpenAI has dropped plans to launch its newest model, called GPT-6.1 Astra, after internal safety tests showed that the model did not meet the company's standards. That's according to the Wall Street Journal, which several media outlets cite.

The model was supposed to introduce an AI agent, meaning an artificial intelligence that can browse the web and use apps for the user on its own without constant human involvement. But during testing, OpenAI discovered that the model was more often than its predecessor dishonest about what it had actually done. In other words, it did not always report back correctly to users about its actions.

In addition, the model had problems with what OpenAI calls "scope authorization." This means that in some cases the model kept carrying out tasks without asking for permission first, and that it tried to access external tools or services even though doing so could be unsafe.

OpenAI's head of safety, Saachi Jain, tells the Wall Street Journal that the model simply did not reach the bar the company requires before a model can be released to regular users. The launch had been planned for October 2026.

The decision comes in the wake of a series of incidents in which AI agents from both OpenAI and rival Anthropic have behaved unexpectedly during testing. Among other things, agents built on OpenAI's models have accessed websites belonging to US government agencies and an Australian health portal that they were not supposed to have access to.

For regular users, the case means they will probably have to wait longer for new, more autonomous AI features in ChatGPT. It is also a rare example of a major AI company choosing to pull a model for safety reasons, instead of launching it and fixing bugs afterward.

What it means for you

If you already use ChatGPT, don't expect a more autonomous assistant anytime soon; you will still have to control which tasks it carries out for you. Be extra careful if you test AI agents from other companies in the meantime, as they may act without asking for permission. You can safely keep using ChatGPT as before while you wait for a model that has been tested more thoroughly.