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

In Matt Wolfe's review, Astra's Computer Use drove Blender directly, generating a humanoid wolf model from a prompt and animating a running cycle.

The author is fed up with benchmark screenshots and "look what app it can build" demos, arguing that people without a monetization system won't make money even with the latest AI.

Linux DO users compared outputs for an SVG pelican-riding-a-bicycle prompt, suspecting Astra was routed to a weaker model, and listed phrasing they use to spot the nerfed version.

A solo founder says GPT-6 Astra finished all of August's bookkeeping hands-off in 20 minutes for about $1.50, versus two hours of their own time or roughly $100 outsourced.

The author tests GPT-6 Astra with the classic SVG pelican-on-a-bicycle prompt, saying stock phrases like "inline SVG + CSS" in the reply suggest the request was routed to a downgraded model.

A developer ran GPT-6 Astra on the HyperFrames harness and rebuilt a16z's Machine Age video as its own launch video using only HTML, and is now running the full code2video benchmark.

thsottiaux briefly replied to a user that Astra keeps building and the team keeps shipping.

Before general release, Astra was the team's biggest internal advantage, and the productivity gain was enough to move some plans up six months to ship at DevDay.

A user says they are rushing to finish remaster work while GPT-6 Astra is fresh, fearing the model will be dumbed down within a month.

A test shows GPT-6 Astra in xhigh mode used 84.6M input tokens (97.1% cached) and 320,946 output tokens to exhaust the weekly limit on the $100 5x plan.

A Zhihu author finds GPT-6 Astra the strongest all-around model in early testing, especially for frontend and 3D generation, but questions the six benchmarks that make up 76% of Artificial Analysis's index, where Astra scored 61 against Fable 5.1's 66.

A benchmark author says early GPT-6 Astra results from VulcanBench are very special and unlike any model behavior seen before, and that he will study the reasoning traces before releasing full results.