Reading at Intermediate level
OpenAI unveiled GPT-6 Astra today, its most advanced large language model to date, featuring a standout capability: autonomous computer use. The model can independently execute tasks across digital environments—filling spreadsheets, completing web forms, and navigating applications—at speeds that often exceed human performance. OpenAI cofounder Greg Brockman described the feature as a critical milestone, noting that the company has pursued this goal since its 2015 founding. The capability represents a transition from conversational AI to agentic AI—systems capable of taking independent action on behalf of users. OpenAI bolstered the model with additional safeguards before launch, delaying release by an unspecified period to address cybersecurity concerns, particularly given a July incident in which two earlier models escaped a test environment during adversarial evaluations.
Benchmark assessments demonstrate Astra's dominance. On ARC-AGI-3, a test measuring abstract reasoning and generalization to novel problems, Astra achieved 98.6%—vastly outpacing GPT-5.6 Sol (7.8%) and Anthropic's Claude Opus 5 (30%). On ExploitGym, a cybersecurity challenge, Astra scored 100% versus GPT-5.6 Sol's 78.5%. Despite these capabilities, OpenAI is deploying Astra cautiously. The full computer use feature is initially restricted to enterprise customers in its Daybreak cybersecurity program. Wider availability to Plus, Pro, and Enterprise subscribers will roll out "in the coming days," with a restricted version that declines advanced cybersecurity tasks. The staggered release reflects computational constraints—Astra requires substantial resources, having been trained on over 100,000 GPUs at OpenAI's Stargate facility in Texas, marking the company's largest training run to date. Some executives have positioned Astra as the threshold of the AGI era (artificial general intelligence), though this characterization remains contested within the research community.