GPT-6 Astra in Practice: Tests, Use Cases, and Feedback

Astra can autonomously operate websites and apps toward a goal and write, run and debug code, with OpenAI reporting a 72.6% score on OSWorld.

A prompt tells Astra to pull Google Analytics, Search Console, GMB and CRM data and decide whether the business is slow because of the market or on its own.

A long post recounts how OpenAI used a new internal model and about 10,000 parallel agents to construct a candidate Navier–Stokes counterexample, and the dispute over whether it preempted mathematicians.

thsottiaux asks in a reply how Astra performs on ultra mode and with /fast, touching on reasoning tiers and speed settings.

thsottiaux says he will see people at the Astra party, a heads-up about an in-person event during the launch period.

Feeding nine rough photos of a studio into GPT-6 Astra, it understood the spatial arrangement, stitched them together and built a fully interactive 3D model of the space.

Apodex scored Astra's scientific discovery ability across six process dimensions, reporting Evidence and Coherence about 20–25% above GPT-5.6 Sol, though gains vary by dimension.

thsottiaux says he is glad prompt injection is being solved across the industry and calls Astra the most capable and most aligned model to date, before trailing off.

The poster calls AA-Omniscience the only benchmark that matters and says passing 95 would be the point where humans hand control to AI.

The author asks whether the Millennium Prize solution came from brute-force compute or a creative "Move 37"-style insight, and speculates on what follows once Astra is wired to robot arms.

The author says GPT-6 Astra is faster, cheaper, more trustworthy and better at almost everything than Fable 5.1, without giving details.

A joking post says that if OpenAI is willing to point Astra at the era's hardest unsolved problems, it might as well solve social media along the way.