What OpenAI's Deployment of GPT-6 Astra Means: Its Computer-Operating Model — September 4, 2026
· CompaniesAutomation
Radar Flash Edition: OpenAI begins deploying GPT-6 Astra, the first model classified by its own framework as having "critical" cyber capability and the best at operating a computer. Specs include 1.05M context, $10/$50 per million, 72.6% on OSWorld 2.0, and what to measure before switching models.
Flash edition. OpenAI has begun deploying GPT-6 Astra, the first model that its own framework classifies as having "critical" cyber capability and the best at operating a computer. Supplement to this morning's Radar.
What happened
The rollout started on Thursday, September 3, for clients of Daybreak, its cybersecurity program, and arrives next week for ChatGPT paid plans and the API (TechCrunch, CNBC). The specs: 1.05 million tokens of context, 128,000 output, and rates of $10 per million input tokens and $50 per million output tokens, with caching at $1 (MarkTechPost). In OSWorld 2.0, the benchmark that measures computer use, it climbs to 72.6%, taking about 40 minutes per task compared to 75 for GPT-5.6 Sol. Greg Brockman, President of OpenAI, concluded with "welcome to the AGI era" (La República). Offensive capabilities—100% in ExploitBench and two unknown vulnerabilities in V8—remain behind the Daybreak Blue list: the Astra that reaches your account does not include them.
Why it matters
Two interpretations. The financial one: $10/$50 is the same rate as Claude Fable 5.1 and thirteen times the token cost of Gemini 3.8 Flash, so per token, it is expensive. But if it takes 47% less time per task and retries less often, the number that matters is not the token price but the cost per completed task. The expectation one: 72.6% in computer use means one in four tasks goes wrong. It is the best in the world at this, and yet it is not an employee: it is a lightning-fast intern whose work must be reviewed.
For your company
Three specific takeaways. One: before migrating anything, measure cost per result, not per token. Take an existing workflow—classified invoices, answered emails—and record euros and minutes per finished unit; without that baseline, comparing prices is just noise. Two: if you are going to give it control of the mouse—opening the CRM, filling out forms, touching spreadsheets—start with a reversible process with automatic result verification; with a nearly 30% failure rate, an irreversible one becomes very costly. Three: enterprise account and minimum permissions. A model that handles applications from the desktop needs credentials that you can revoke at three in the morning, and a log of what it opened and with which key.
Frequently Asked Questions
Can I use GPT-6 Astra in my company now?
With ChatGPT Plus, Pro, Business, or Enterprise, next week; via API, as well. What is not arriving is the offensive cybersecurity part, reserved for verified organizations.
Is it dangerous for a model to find vulnerabilities on its own?
What matters to you is not the model, but the timeline: if a machine discovers and exploits unknown flaws, the window between patch and attack is measured in hours. Keep an inventory of what is exposed to the internet and set patching deadlines in writing.
Is it worth switching models now?
Only if the bottleneck is the model's capability. If the problem is that no one has defined when a task is done correctly, switching models won't fix anything.