OpenAI's new GPT-6 Astra can operate a computer like a human, company says
OpenAI has released GPT-6 Astra, its most advanced model, which can operate a computer by navigating interfaces and completing tasks. The company says it outperforms previous models and rivals on benchmarks, including a 98.6% score on ARC-AGI-3. OpenAI president Greg Brockman described the launch as the beginning of the AGI era.
OpenAI’s GPT-6 Astra introduces “computer use,” letting the model navigate interfaces, fill forms, and process spreadsheets at speeds the company claims often exceed human capability. The model scored 98.6% on ARC-AGI-3, a reasoning benchmark, and 100% on ExploitGym, a cybersecurity test—far outpacing its predecessor GPT-5.6 Sol. However, Astra’s initial rollout is limited to select enterprise clients in OpenAI’s Daybreak cybersecurity program, with broader availability for Plus, Pro, and Enterprise users expected soon. The company cites the model’s unprecedented scale—trained on over 100,000 GPUs at its Texas Stargate site—as a reason for the phased release. OpenAI also notes Astra will refuse advanced cybersecurity tasks for general users, a safeguard added after a July incident involving earlier models.
Widespread adoption of computer-operating AI could reshape office work, automating routine digital tasks and potentially displacing some administrative roles. Yet the technology may also introduce new risks, such as over-reliance on automated systems for sensitive actions like financial transactions or data handling. While OpenAI frames this as the dawn of an “AGI era,” the limited enterprise-first release suggests cautious deployment, acknowledging that societal readiness and safety guardrails remain works in progress.