The San Francisco-based company began rolling out the model Thursday, initially to a limited group of organizations. OpenAI said GPT-6 Astra will be made available in the coming days to ChatGPT Plus, Pro, Business and Enterprise subscribers, as well as through its API and Amazon Web Services.
Astra is designed to handle computer-based tasks more autonomously, including filling out online forms, updating customer records, organizing calendars and conducting research. OpenAI also says the model can create documents, spreadsheets and presentations.
Major gains in reasoning and tool use
OpenAI’s published benchmarks show significant improvements over its previous models, particularly on tasks requiring extended reasoning and the use of external tools.
Read also : OpenAI tightens AI safety controls after agents escape security test
On ARC-AGI-3, a test focused on advanced reasoning, GPT-6 Astra scored 98.6%, compared with 7.8% for GPT-5.6 Sol and 39.2% for Claude Opus 5, according to figures cited by Numerama.
The model also recorded 97.6% on FrontierMath Tier 4 and 41.4% on AutomationBench, compared with 83% and 18.1%, respectively, for GPT-5.6 Sol.
Astra’s performance was also stronger on tasks involving real computer environments. On OSWorld, which measures an AI model’s ability to operate software, OpenAI reported a score of 72.6%, compared with 65.7% for GPT-5.6 Sol. The company said the average time required to complete a task fell from 75 minutes to 40 minutes.
The gains are less pronounced on some pure programming benchmarks. Astra scored 74.1% on DeepSWE v1.1, compared with 70.8% for GPT-5.6 Sol, showing that the new model’s biggest improvements are concentrated in longer tasks, autonomous tool use and scientific work.
Cybersecurity becomes a key focus
OpenAI is also highlighting a major increase in Astra’s cybersecurity capabilities.
Read also : OpenAI expands review after identifying additional AI agent containment breaches
The company said the model has reached the “Critical” threshold for cybersecurity under its Preparedness Framework. According to OpenAI, Astra can autonomously identify and exploit vulnerabilities in heavily protected computer systems.
The model’s most sensitive cybersecurity capabilities will be restricted to specialized cyberdefense users, while automated monitoring can interrupt tasks when necessary, including tasks unrelated to cybersecurity.
The announcement comes after a summer marked by growing concerns over the ability of advanced AI agents to carry out cyberattacks with limited human intervention. OpenAI said Astra was not involved in a previously reported incident in which OpenAI-built agents accessed another company’s testing environment and the open-source platform Hugging Face.
Astra rollout starts with selected users
GPT-6 Astra is initially being made available to a limited number of organizations through OpenAI’s Daybreak access program.
Read also : More than 1,200 AI workers urge US government to “deliberately pace” development
The company plans to expand access to ChatGPT Plus, Pro, Business and Enterprise users, as well as developers through the API, in the coming days.
OpenAI says Astra can complete individual tasks at a lower cost than its predecessor because it can use fewer tokens and finish some tasks more quickly. However, its API pricing remains among the higher rates in the current AI market.
