Sam Altman says Astra model is so powerful, OpenAI can't launch it now

OpenAI CEO Sam Altman has announced that Astra is a powerful AI model, and the startup wants to release it, but it can't at least right now. Instead, OpenAI is working on making Astra safer for a public release

In Short

• Sam Altman says Astra AI model is too powerful to release
• OpenAI says it is stronger than GPT-5.6 Sol
• OpenAI is working on making Astra safer

Frontier AI development is showing no signs of slowing down. Just weeks after OpenAI launched GPT-5.6 Sol, it now has an even more powerful AI model called Astra. But OpenAI CEO Sam Altman has now admitted that Astra is too powerful to be released. Instead, the AI startup needs to first make sure that the model is safe before it is given to the public.

On X, Sam Altman declared that OpenAI had no plans to restrict Astra. “Astra is a powerful model and we are working to make it generally available,” he said. The OpenAI chief added, “We do not think it is a good strategy to keep powerful models to a chosen few.”

But according to Altman, the startup needed to work on ensuring that Astra would be safe before its public release. He said, “Given its cyber capabilities, we need a little big longer to do do this safely. but hopefully not too long!”

Altman’s comments come at a time when there is growing debate over access to frontier AI models. Ever since the US government temporarily banned foreigners from using Anthropic’s Mythos and Fable models, there are fears that such restricted access may leave powerful models to the hands of a few.

The US government has also considered blocking Chinese open-weight models. Though several companies, including Nvidia, Microsoft, and Amazon, have urged the White House to not put such a ban.

OpenAI is working on making Astra safer

In a blog post, OpenAI said its latest internal evaluations of Astra showed significant advances in agentic coding and cybersecurity, and that it could not rule out the model reaching a critical level of cyber capability.

Keep in mind that OpenAI’s GPT-5.6 Sol and Anthropic’s Mythos are already capable of finding cybersecurity flaws that were missed by humans. The same models have also been involved in breaches during testing in the past.

OpenAI said that it was testing Astra as part of its Preparedness Framework that was first launched in December 2023 as a guide to identify capability progress and decide what steps to take as such capabilities emerged.

Under that framework, OpenAI says that a model reaches the Critical cybersecurity threshold if it can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention. Or if the AI model can devise and execute end-to-end novel strategies for cyberattacks against hardened targets.

What do we know about Astra?

OpenAI said its preliminary evaluations of Astra were strong enough that it could not rule out that level at this stage. Though it has clarified that Astra was not involved in the recent cyberattack on Hugging Face by rogue OpenAI agents.

The startup’s previous models, including GPT-5.6 Sol, had been assessed at the High threshold rather than the Critical threshold for frontier cyber capabilities.

As Astra shows more progress, OpenAI has scaled up robustness testing of safeguards and security controls to match a possible deployment of these capabilities. The company also said it was pausing internal activities involving Astra that did not yet meet its stricter safety requirements.

The discussion around Astra follows another OpenAI blog post where the company claimed that the AI model had resolved or made progress in ten of the hardest math problems out there. This included coding theory, operator algebras, quantum complexity, and lattice cryptography. Later, an Anthropic engineer claimed that the Fable 5 AI model could also solve five of those problems.

OpenAI has not disclosed whether Astra would be part of the GPT-5.6 family, or be the start of GPT-6. While Astra is not accessible to the general public, as per reports, OpenAI CEO Sam Altman has showcased the model to federal officials already.

- Ends