What it means when AI can operate your computer on its own

OpenAI shipped GPT-6 Astra on September 3. It is not just a smarter chatbot. It is built to operate a computer on its own — clicking, typing, navigating apps — without anyone guiding each step. OpenAI calls it their most intelligent and aligned model. But the more useful description is simpler: it can get things done on a screen by itself.

The benchmarks show why this matters. On OSWorld 2.0 — a test that measures how well an AI completes real desktop tasks — Astra scored 72.6% and finished those tasks 47% faster than the previous version. On ARC-AGI-3, a general problem-solving test, it scored 99.9%. These numbers cross a threshold where automated computer work becomes practical, not just impressive in a demo.

There is a deliberate caution built in. Astra scored 100% on ExploitBench, a cybersecurity test. OpenAI locked the most sensitive features behind a special access program, rating this the first model it considers a "Critical" cybersecurity risk — because it can find unknown security flaws in protected systems on its own. The power that makes it useful also demands careful handling.

The price reflects the ambition. Astra costs $10 per million input tokens and $50 per million output tokens — roughly 2.5 times the cost of the previous flagship. It uses fewer steps to complete a task, so the total cost per job may still be reasonable. But this is a model built for high-value, complex work — not for cheap, high-volume tasks.

The practical shift is worth acting on now. If you have a repeatable, computer-based task — moving data between systems, running a research workflow, filling structured reports — Astra is strong enough to test in a real process, not just a demo. The question is no longer whether AI can do it. The question is where it fits and what it costs.