It's September 3, 2026 and OpenAI has just officially announced its newest flagship model. Dubbed “GPT-6 Astra," it's being described as its most intelligent and aligned model yet. It is the first upgrade since the launch of GPT-5 a little over a year ago. While celebratory tones are in abundance from OpenAI and many of its followers, a more tempered warning is being stressed:
OpenAI has detailed ground breaking benchmark numbers, but the model's features are advanced enough to warrant new forms of safeguards prior to its general release.
OpenAI says that Astra draws on years of research and large gambles on pretraining, reinforcement learning, and alignment work and that the system is state-of-the-art in terms of computer use, web browsing, software engineering, cybersecurity, scientific reasoning, and general professional work.
Its benchmark numbers are impressive:
Asta reportedly saturates FrontierMath Tier 4, with a 98% score, after reportedly using this benchmark to help already settle longstanding unsolved problems in mathematics, it has exceeded the human action-efficiency baseline on 96% of levels onARC-AGI-3, the benchmark for actual reasoning not memorization, nearly equaling human performance; it reportedly earned a perfect score on ExploitBench, a benchmark measuring usefulness in the realm of security vulnerability identification.
President of the company Greg Brockman calls the release not an evolutionary advance but "a genuine step toward turning over complex cognitive tasks to AI systems."
Apart from these quantitative measures, OpenAI highlights Astra's gains in understanding user intent. The organization explains that it will stay 'oriented' in complex tasks, abide by constraints, and make practical decisions in ambiguously framed instructions by inferring information from the context rather than guessing or over-requesting. OpenAI is introducing a new test to evaluate the systems' 'respect for intended use limits,' specifically after a past Hugging Face-related incident. The new model, as per OpenAI, should not overreach or assume more about a task when dealing with very difficult inputs.
Astra is also presented as both an engineering and creative tool.
The feature called sites will permit ChatGPT to produce, display, and distribute websites, applications, and games from a single prompt, using approximately 20% fewer tokens in complex creative tasks than previous models.
The difference from past releases isn't just about what Astra can do; it's about how OpenAI plans to release it. The company's revealed Astra to be the first of their models to hit the threshold they consider 'critical' for cybersecurity risks according to their own Preparedness Framework—the classification scheme used to describe how dangerous a model's features could become in the wrong hands.
As such, it's not being released more widely immediately. It's initially being rolled out to a closed set of organizations currently in OpenAI's Daybreak application-based cybersecurity program. It's only over the coming days that the wider set of ChatGPT Plus, Pro, Business, and Enterprise subscribers are set to be let loose as well as giving it out on the OpenAI API and through Microsoft Azure and AWS.
The tradeoffs are also public knowledge—and not ones OpenAI has kept quiet about. As stated above, the safeguards have the potential to flag legitimate use cases, including defensive cyber work or long-standing autonomous runs, as a bad actor action that results in slowed, halted, or suspended use on both ChatGPT and Codax and requires the user to verify each suspected action individually.
It’s a risk OpenAI has accepted for a model they've personally identified as higher risk than its previous, allowing the sacrifice of a little usability in return for a better check on that risk.
Details on the price have put the API use of Astra at around $10 per million input tokens and $50 per million out, with adjustable levels for reason-effort levels going up to the highest in order to solve truly hard problems. ChatGPT subscribers will gain a certain allocation of this on current plans, with usage credits being available for extra use, while those with subscriptions to the Pro, Business, and Enterprise plans will also be given an additional subscription tier named 'Astra Pro.'
