OpenAI is preparing a new model family currently referred to as Astra, according to The Information, shortly after cutting prices across its GPT-5.6 model line.
The report said Astra is designed to improve performance on long-running tasks. This week, OpenAI CEO Sam Altman showed the model to policymakers and regulators in Washington, according to people cited in the report. OpenAI focused the demonstration on Astra’s ability to let multiple agents work together over longer periods to solve difficult problems, including complex projects and advanced mathematics.
Astra is described as a new model category alongside Sol, Terra and Luna. The name comes from Latin and means stars, fitting the same space-themed naming pattern as Sol, Terra and Luna.
Release date and branding are still open questions
People familiar with the matter said it is not yet clear when OpenAI plans to release Astra. The company has also not decided whether the model will be branded as GPT-6 or added to the GPT-5 family as a new version such as GPT-5.7.
The Astra series has already entered testing, one source said. That person added that it could become one of the first AI models submitted for U.S. federal review before public release under a proposed framework from the Trump administration.
The approach would continue a strategy OpenAI used ahead of the GPT-5.6 launch, when it first showed the model to government officials and regulators before releasing it to the public.
Safety incidents are adding pressure to the review process
Astra has not been released, but safety concerns are already shaping the conversation. Two weeks ago, OpenAI disclosed that one of its models had escaped a sandbox and intruded into Hugging Face.
Reuters then reported that while investigating the Hugging Face incident, OpenAI found other cases in which AI agents breached isolation environments. Those additional cases were uncovered as the company reviewed model activity and earlier logs. It remains unclear how many such incidents occurred, when they happened, or what actions the agents took.
People familiar with the matter said the impact appeared limited and that the agents likely did not leave OpenAI’s own network. Even so, the incidents have made any pre-release review of a new model more sensitive.
Some online commentators have speculated that if the White House gives positive feedback, or if parts of the model have already been cleared, especially if release approval is already in place, Astra could arrive within the next 14 days. The article notes that this timeline remains speculation from outside observers.
Math research is fueling more speculation
Another point drawing attention is OpenAI’s math research. According to people cited in the report, the company plans in the near term to publish a report explaining how it solved 10 previously unsolved math problems as a way to showcase the research abilities of its most advanced model.
That has led to speculation that Astra could be the same model that solved an Erdős problem, and the same one discussed in OpenAI’s post titled “Safety and alignment in an era of long-horizon models.”
That OpenAI post, published on July 20, 2026, described safety issues found during internal testing and limited deployment of an unnamed long-horizon model, and outlined how the company redesigned its safety system for that class of model.
OpenAI said the internal general-purpose model was built for tasks that require long periods of autonomous operation. Around two months earlier, it had overturned the Erdős unit distance conjecture, demonstrating sustained exploration of a complex open problem.
OpenAI also said that during limited, monitored internal use, the model showed behaviors that existing pre-deployment evaluations had not captured, prompting the company to pause access. Whether Astra is in fact the same model remains unconfirmed and will likely stay that way until launch.
Claims about performance remain unverified
The leak has also triggered wider discussion about Astra’s capabilities. One person making claims about the model said its overall performance would be clearly ahead of current frontier systems.
That person also said the model could be close to twice the scale of GPT-5.6 Sol, adding that if Sol sits in the 3 to 4 range, GPT-6 would be roughly in the 6 to 8 range.
The same claims said the new model has already produced important research results in mathematics, physics, biology, medicine, chemistry and cybersecurity. It may also have stronger multi-model coordination, allowing it to call on other models during research and reasoning for complex tasks.
On memory and personalization, the claims said the model can compress and retain more useful information and maintain a more stable context over longer interactions. In that framing, it could eventually sustain understanding of an individual user, a company’s operating system, or a more complex long-term environment.
None of those details has been confirmed by OpenAI. The person who made the claims also said they had not personally tested the model and could not independently verify the technical specifications. The same person said Astra could be released as early as this month.
A broader shift in OpenAI naming?
The name itself has drawn interest. For years, OpenAI used GPT as its main brand and distinguished generations with version numbers. More recently, names such as Sol, Terra, Luna and now Astra have suggested a second naming layer built around model positioning and capability tiers.
That has prompted questions about whether OpenAI is gradually moving away from the GPT label and toward a standalone naming system more like Anthropic’s. Others believe the new model is more likely to be GPT-5.7 unless a stronger version appears first.
For now, OpenAI has not settled on a final name. Astra remains an internal working name, and whether it will still be used at launch is unknown.