OpenAI says GPT-6 Astra scored 72.6% on OSWorld 2.0 at about 40 minutes per task
In brief
- GPT-6 Astra, OpenAI's model for computer use across apps and interfaces, launched September 3, 2026.
- Astra scored 72.6% on OSWorld 2.0 versus 65.7% for GPT-5.6 Sol, per Crypto Briefing.
- Parallel agents reached ChatGPT Pro subscribers in a mid-October update.
- LeoDex News couldn't independently verify OpenAI's benchmark figures.
What the benchmarks show
OSWorld 2.0 is a benchmark that tests how well an AI completes real tasks on a desktop operating system. Astra's 72.6% compares with 65.7% for its predecessor, GPT-5.6 Sol, which took around 75 minutes per task on the same test (Sol had been the previous benchmark for OpenAI's agentic work).
That's a big cut in time per task.
On Agents' Last Exam, a benchmark built around professional-grade tasks, Astra scored 59.3% versus 53.6% for GPT-5.6 Sol. Its ScreenSpot-Pro score climbed to 92.7%, up from 76.9%. ScreenSpot-Pro measures how reliably a model can locate and identify on-screen elements (buttons, menus and similar) in professional software.
All of these figures are OpenAI's, as reported by Crypto Briefing. No OpenAI document was linked, and LeoDex News hasn't verified them independently.
Crypto Briefing also passed along several claims it described only as "reportedly" true, including a 99.9% score on the ARC-AGI-3 reasoning benchmark and Excel financial modeling about four times faster than the 2023 Excel Financial Modeling World Cup winner. It didn't cite a source for those, so we're treating them as unconfirmed.
What's changed since launch
Astra isn't a new release. It's been out since early September, and the latest development came in a mid-October 2026 update that brought parallel agents and tool-use improvements to ChatGPT Pro subscribers. Parallel agents let multiple AI workers tackle different parts of a job at the same time.
Why human oversight still matters
Astra's focus is computer use: clicking through interfaces, moving between applications and finishing multi-step jobs without a human steering every move. So how close is "close to expert human speed"? That's OpenAI's framing, not a settled fact.
Crypto Briefing's own read is more cautious. It said a 72.6% success rate means the agent still falls short on a meaningful share of tasks, so human oversight remains necessary.
Frequently asked questions
What is OSWorld 2.0 and how did GPT-6 Astra score on it?
OSWorld 2.0 is a benchmark that tests how well an AI completes real tasks on a desktop operating system. According to OpenAI figures reported by Crypto Briefing, Astra scored 72.6% at roughly 40 minutes per task. GPT-5.6 Sol scored 65.7% at around 75 minutes per task.
Does GPT-6 Astra remove the need for human oversight?
No. Crypto Briefing said a 72.6% success rate means the agent still falls short on a meaningful share of tasks, so human oversight remains necessary. OpenAI's team said its agents are close to expert human speed, but that's the company's own framing.
What did the mid-October 2026 update add?
A mid-October 2026 update brought parallel agents and tool-use improvements to ChatGPT Pro subscribers. Parallel agents let multiple AI workers handle different parts of a job at the same time.
