OpenAI's GPT-6 Astra Nearly Matches Expert Humans at Using Computers

OpenAI's newest agent brain is out, and it moves fast. GPT-6 Astra launched September 3, 2026, built for computer use: clicking through interfaces, hopping between apps, and finishing multi-step jobs without a human babysitting every click. The pitch from OpenAI's team is blunt — these agents are nearing expert human speed.

The scoreboard backs it up. On OSWorld 2.0, the benchmark that tests real desktop tasks, Astra hit 72.6% accuracy at roughly 40 minutes per task. Its predecessor, GPT-5.6 Sol, managed 65.7% — and took about 75 minutes per task. More accurate, nearly half the time. That combo is what matters when you're paying for agent hours.

Other benchmarks moved too. Agents' Last Exam, a professional-grade gauntlet, saw Astra score 59.3% against Sol's 53.6% — and it reportedly burned 65% fewer tokens than some Claude evaluations on that exam. Tokens are the units of text a model chews through, so fewer of them generally means cheaper, quicker answers. On ScreenSpot-Pro, which measures how reliably a model spots buttons, menus and fields in pro software, Astra climbed to 92.7% from 76.9%.

Then there's the flashy one. Astra reportedly finishes tasks in Excel Financial Modeling about four times faster than the 2023 winner of the Excel Financial Modeling World Cup — a contest for the planet's elite spreadsheet builders. It has also reportedly saturated the ARC-AGI-3 reasoning benchmark at 99.9%, with safety metrics improving too, including a reportedly flawless record in certain computer-usage scenarios.

OpenAI hasn't stopped tinkering. A mid-October 2026 update rolled out parallel agents and better tool use for ChatGPT Pro subscribers, letting multiple AI workers split up a job instead of grinding through it one step at a time. Closer to handing a project to a small team than to a single assistant.

Here's the catch. A 72.6% success rate still means the agent whiffs on a meaningful slice of tasks, so humans stay in the loop for now. And the gap between 'close to expert humans' and 'faster than expert humans' is exactly where the money question lives. OpenAI frames its agents as nearing expert speed overall — but on narrow tasks like Excel modeling, the model is already pulling ahead.