Three numbers in OpenAI's official GPT-6 Astra video: 47%, 40 minutes, 1.9x
A breakdown of OpenAI's official demo: 72.6% on OSWorld 2.0 at about 40 minutes per task, roughly 47% faster than GPT-5.6 Sol, and 1.9x faster on Mind2Web with the Codex harness.
OPENAI JUST SHARED A GPT-6 ASTRA VIDEO, AND 3 NUMBERS INSIDE IT MATTER MORE THAN THE MODEL ANNOUNCEMENT: 47%, 40 MINUTES, 1.9X don’t watch it like another AI demo. watch what Astra is actually doing on the computer. OpenAI reports 72.6% on OSWorld 2.0 at roughly 40 minutes per task. GPT-5.6 Sol needed roughly 75 minutes while scoring 65.7%. that’s about 47% less time per task. then OpenAI changed the Codex harness around it. combined with Astra, task completion reached 1.9x the speed of the previous GPT-5.6 Sol experience on Mind2Web. and look at what the model is being tested to actually do: → navigate software → fill forms → update CRM records → research across the web → build websites → run frontend QA → install and test software this is why the official video matters. OpenAI isn’t showing a model sitting inside a better chat box. they’re showing the direction of the product: give the model a computer, give it a finish line, then measure how much work disappears before you touch the keyboard again.
This post quotes another post: x.com/lummox_eth/status/2100288151897981255