OpenAI’s new model isn’t built to answer your question. It’s built to use your computer. VentureBeat reported the release of GPT-6 Astra on September 3, and OpenAI’s own line is that it’s “the world’s best computer use model.” Instead of replying in a chat window, Astra works software the way a person does: fills out forms, updates CRM records, organizes calendars, does research, works in spreadsheets and Power BI, troubleshoots other programs. On the OSWorld 2.0 benchmark, OpenAI’s numbers, it finished 72.6 percent of tasks, up from 65.7 for its predecessor, in about 40 minutes a task instead of 75.
It’s also the first OpenAI model to reach the “Critical” tier of the company’s own cybersecurity framework, which means it can find unknown vulnerabilities and build exploit chains on its own; during evaluation it found two real ones, which OpenAI says it disclosed to the maintainers. OpenAI says it paused frontier training for roughly two weeks after the Hugging Face breach in July and tightened its infrastructure and monitoring before this release.
Enterprise customers got it first through a program OpenAI calls Daybreak; ChatGPT Plus, Pro, Business and Enterprise, the API, AWS and Azure follow over the coming days. Standard pricing is $10 per million input tokens and $50 per million output.