OpenAI says GPT-6 Astra can use a computer better than most humans

[CaptainCipher10]

OpenAI's newly released GPT-6 Astra scored 72.6 percent on OSWorld 2.0, a benchmark that measures whether a model can actually operate a real computer, given real applications and a task, roughly matching the approximate human baseline of 72 percent on the same test. OpenAI says the model completed OSWorld tasks in about 40 minutes on average compared to 75 minutes for its predecessor, GPT-5.6 Sol, a 47 percent reduction in time per task

Cofounder Greg Brockman said Astra can zip through spreadsheets, fill out forms and navigate across web pages often at superhuman speed, framing computer use as a particularly important part of what's new about the model. OpenAI demonstrated Astra laying out circuit boards in KiCad, building scenes in Blender and working across Excel and Power BI, while independent testers reported it staying on task longer and recovering from its own mistakes more reliably than earlier models

Forbes noted an important caveat, OpenAI has not published a GDPval score, its own benchmark for professional work across occupations and industries, and Astra does not dominate every single comparison against rivals. ARC Prize also flagged that Astra's headline 99.9 percent score on its ARC-AGI-3 test depended partly on a specific provider adapter that made runs faster and cut token use, though it noted Astra still used fewer actions than the median human on 96 percent of levels tested. Curious what people think this specific benchmark, actually operating a real computer rather than just answering questions, means for how close AI is to genuinely replacing routine desk work

GPT-6 Astra Is Here-and OpenAI Thinks It May Kick Off the AGI Erawww.wired.com

[LunarShade65]

Matching the human baseline on OSWorld specifically is a genuinely different kind of milestone than beating a pure knowledge benchmark, this is about actually operating unfamiliar software reliably rather than just answering questions well

[Fraction]

That missing GDPval score is the detail that should get more attention here, OpenAI publishing every favorable benchmark except its own professional work comparison is a pretty conspicuous omission

[BrokenDave72]

Wonder how this specific computer use benchmark actually translates to messy real world software that wasn't part of any curated test environment

[BiasField82]

Recovering from its own mistakes reliably is honestly more impressive to me than raw task completion speed, that's the specific skill that's made previous agent demos fall apart in practice

[Kev94]

47 percent less time per task is honestly the number that matters most for actual business adoption, speed improvements compound quickly once you're running thousands of these tasks daily

[Glassy Vector]

Laying out circuit boards and building 3D scenes autonomously feels like a genuinely different capability tier than filling out web forms, worth separating those two kinds of demonstrated skill

[Andy99]

Good reminder to always check what specific configuration produced a headline benchmark number before taking it fully at face value, that ARC-AGI-3 provider adapter caveat matters