OpenAI / Wikimedia Commons (Public domain)
OpenAI's team says its agents are closing in on expert humans at operating software, and the benchmark numbers show big gains in both speed and accuracy
OpenAI’s team says its computer-using agents are now close to matching expert humans on speed. That claim arrives with GPT-6 Astra, a model built to operate software the way a person at a keyboard would.
What Astra actually does #
GPT-6 Astra launched on September 3, 2026. Its focus is computer use: clicking through interfaces, moving between applications, and finishing multi-step jobs without a human steering every move.
The headline number comes from OSWorld 2.0, a benchmark that tests how well an AI completes real tasks on a desktop operating system. Astra scored 72.6% at roughly 40 minutes per task.
Its predecessor, GPT-5.6 Sol, scored 65.7% on the same test at around 75 minutes per task. Astra is more accurate and takes close to half the time per task, which is the combination that matters for anyone paying for agent hours.
The benchmark scorecard #
OSWorld is not the only test where Astra posted gains. On Agents’ Last Exam, a benchmark built around professional-grade tasks, Astra scored 59.3% compared with 53.6% for GPT-5.6 Sol.
Astra also reportedly used 65% fewer tokens than some Claude evaluations on that exam. Tokens are the units of text a model processes, and fewer tokens generally means lower costs and quicker answers.
The model’s ScreenSpot-Pro score climbed to 92.7%, up from 76.9% for GPT-5.6 Sol. ScreenSpot-Pro measures how reliably a model can locate and identify elements on a screen, such as buttons, menus, and fields in professional software.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The flashiest comparison involves spreadsheets. Astra reportedly completes tasks in Excel Financial Modeling about four times faster than the 2023 winner of the Excel Financial Modeling World Cup, a competition for elite spreadsheet builders.
Astra has also reportedly saturated the ARC-AGI-3 reasoning benchmark at 99.9%. Safety metrics improved as well, with a reportedly flawless safety record in certain computer usage scenarios.
Background: from Sol to Astra #
GPT-5.6 Sol was the previous benchmark for OpenAI’s agentic work. Astra’s improvements over it come in lower time-to-task-completion and higher accuracy, the two metrics that define whether an agent is useful or merely interesting.
OpenAI has kept building on the release. A mid-October 2026 update introduced parallel agents and improvements to tool use for ChatGPT Pro subscribers.
Parallel agents let multiple AI workers tackle different parts of a job at the same time. Instead of one assistant doing tasks in sequence, it is closer to handing a project to a small team that divides the work.
What this means #
The gap between “close to expert humans” and “faster than expert humans” is where the economic story lives. OpenAI’s team frames its agents as nearing expert speed overall, while specific tasks like Excel modeling already show the model pulling ahead.
A 72.6% success rate still means the agent falls short on a meaningful share of tasks, so human oversight remains part of the picture for now.
The token efficiency comparison against Claude evaluations signals that cost per task is becoming a battleground, not just raw benchmark scores.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our