Three days after telling Congress it was building the off switch, OpenAI shipped the thing the off switch is for. GPT-6 Astra — the computer-use model that crossed the Preparedness Framework’s Critical threshold earlier this week in sandboxed tests — began its general rollout on September 3, reaching paid ChatGPT tiers, the API, Azure, and Bedrock “in the coming days.” The sandbox is empty. The operator is at large.
The benchmark table does the arguing. Astra scores 72.6 percent on OSWorld 2.0, the standard test of driving a real computer, against 65.7 for GPT-5.6 Sol and 70.2 for Claude Opus 5. Screen grounding — finding the right pixel to click — hits 92.7 percent. Agents’ Last Exam, the broad agentic gauntlet, goes to 59.3. Greg Brockman called it “the start of AGI,” which is a marketing sentence wrapped around a real fact: the model that operates the computer is now a product you can subscribe to.
The caged evaluation of Monday is the general release of Friday — the fastest a frontier threshold has ever become a subscription tier.
The Cyber Cohort Gets It First
The rollout order is the story inside the story. The first customers are organizations in OpenAI’s Daybreak cybersecurity program — the same defensive community that spent the week reading OpenAI’s letter about automated shutdown capabilities. The general version ships with the sharp edges disclosed: it refuses advanced offensive cyber tasks, the company says, the refusal list being the difference between the caged build and the shipped one. A Pro variant rides along for Business and Enterprise plans, at reported pricing of $10 per million input tokens and $50 per million output.
Read the week as one sentence. Monday: a model crosses the threshold where it can chain zero-days, and OpenAI keeps it caged. Wednesday: a letter to lawmakers promises automated shutdowns, tool-step monitoring, and harder isolation. Friday: the model ships to everyone, minus the dangerous tasks, with the governance apparatus announced before the product. Whatever else this is, it is not the old release rhythm — capability first, policy later. The policy is now part of the launch sequence.
The Takeaways
- GPT-6 Astra began general rollout September 3 to ChatGPT Plus/Pro/Business/Enterprise, the API, Azure, and Bedrock, after a week in caged evaluation.
- Benchmarks: 72.6% OSWorld 2.0 (vs 65.7% GPT-5.6 Sol, 70.2% Claude Opus 5), 92.7% ScreenSpot-Pro, 59.3% Agents’ Last Exam.
- Daybreak cybersecurity-program organizations get first access; the general version refuses advanced offensive cyber tasks.
- The rollout order is governance-led: threshold disclosure, shutdown-capability letter, then general release — policy before product.
- Greg Brockman framed it as “the start of AGI”; the verifiable fact is that a frontier computer-use agent is now a subscription product.
