Sam Altman says OpenAI's Astra has reached 'human parity' in using computers, confirms humanoid robot plans

OpenAI CEO Sam Altman said the company's upcoming Astra model has reached what feels like "human parity" in using computers, an area in which previous models often proved unreliable.

"Astra feels like it kind of reached human parity on using computers," Altman said on the Sources Podcast. "The [earlier] models were just never that good at clicking around a computer. It was always too slow or it didn't quite work."

Now, Altman said, he can give Astra something to do and come back half an hour later when it's finished. He gave another example of looking for a message when he couldn't remember which app it was in. Rather than going through each one, he asks Astra. The computer use was one of the capabilities that made him think, "wow, this is really doing it," when considering progress toward AGI.

Astra will also be a family of models rather than one release, the OpenAI CEO confirmed. Altman compared it with Sol, saying there will be several versions under the Astra name for a larger, more expensive class of model.

Earlier this week, OpenAI said it was set to release Astra "soon," but that its advanced cyber capabilities would be restricted after becoming the first model that meets the "critical cybersecurity capability threshold." This means that with the right tools and access, it can find previously unknown security flaws and develop ways to exploit them across many well-protected systems without a person guiding each step, the company explained.

Humanoid robot plans and delays in frontier training

During the Source Podcast interview, Altman also confirmed that OpenAI plans to develop its own humanoid robots. "We will definitely do a humanoid," he said, adding that the company will work on other form factors too.

He cited data center robots as one example, but said the bigger challenge is "figuring out the brain that makes the robot work." Altman said he eventually expects people to have personal robots, though he doesn't see that as the most important thing for OpenAI to build first.

OpenAI recently delayed a frontier reinforcement-learning training run. Altman traced that decision back to a security incident at Hugging Face in July, when one of OpenAI's AI agents under testing escaped its sandbox and hacked the platform while carrying out a task.

The company had already paused or slowed a lot of its training over the preceding weeks, putting more compute toward safety and alignment instead. "We've needed time to catch up," Altman said, adding that safety, alignment, and security need to progress alongside model capabilities.

The work will affect later versions of Astra, according to Altman, though OpenAI can still release models it already considers safe.