OpenAI has begun rolling out GPT-6 Astra, the company's most capable model to date, according to an announcement on OpenAI's website. Access started with a limited group of organizations enrolled in Daybreak, OpenAI's cybersecurity program, and the company says it will reach ChatGPT Plus, Pro, Business and Enterprise subscribers — its personal and corporate paid tiers — in the coming days, along with OpenAI's API, Microsoft Azure and AWS Bedrock.
Hipertextual and Mobile Time both reported the same staged rollout in their coverage of the launch. Daybreak, the program getting first access, is OpenAI's early-access track for cybersecurity clients, according to Hipertextual.
A model built to operate a computer
The industry calls this category "computer use": instead of only answering questions in text, the model sees the screen and acts on it directly, clicking, typing and moving between programs the way a person would. OpenAI says Astra was trained to operate applications and browsers directly, rather than only telling a user what steps to take.
Among the examples the company lists — and that Mobile Time also reported — are filling out online forms, updating records in CRM software (the tools companies use to track customer interactions), organizing calendars, researching topics online and running automated software tests.
OpenAI also says the model handles ambiguous instructions differently: when information is missing, Astra fills the gap using available context on its own, and only asks the user when the answer could materially change the outcome. Mobile Time described the same behavior in its coverage of the launch. In practice, that means relying less on detailed instructions for routine tasks, such as organizing a schedule or filling out a record, and reserving human confirmation for decisions that actually change the outcome of the work.
Benchmark scores
According to figures published by OpenAI and confirmed by Mobile Time, GPT-6 Astra scored 98% on FrontierMath Tier 4, a set of advanced math problems that typically takes human mathematicians weeks to solve. The model also reached 99.9% on ARC-AGI-3, a benchmark for learning new tasks without task-specific training, and a perfect 100% on ExploitBench, which measures a model's ability to find security vulnerabilities in software.
GPT-6 Astra succeeds GPT-5.6 Sol in OpenAI's model lineup, and it's against that earlier version that the company measures the performance gains it is announcing.
On OSWorld 2.0, a benchmark for autonomous computer use, OpenAI reports a 72.6% score at roughly 40 minutes per task. Mobile Time recorded a similar figure, near 73%, with tasks completed 35 minutes faster than with the prior model. Both sources cite GPT-5.6 Sol, OpenAI's previous flagship, which scored 65.7% at roughly 75 minutes per task, as the comparison point.
API pricing
A token is the unit of text a language model processes — on average, a piece smaller than a whole word. Developers who access GPT-6 Astra through OpenAI's API — the channel companies use to plug the model into their own products, separate from ChatGPT — will pay $10 per million input tokens and $50 per million output tokens in standard mode, according to figures confirmed by Hipertextual and Mobile Time.
Both outlets also mention a pricier tier for heavier use, but they give different figures for it, so those numbers are left out of this report. Pricing through the API is separate from what individual subscribers pay for ChatGPT Plus, Pro, Business or Enterprise, which give access to the model through the chat interface at a flat monthly cost rather than by usage.
For now, GPT-6 Astra remains limited to the initial group of Daybreak organizations. OpenAI has not announced a date for extending access to free ChatGPT users.
