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

A video shows GPT-6 Astra also performing impressively on the RoboDojo robotic manipulation benchmark.

The author calls Astra the best computer-use model ever, with its benchmark score rising from 77 to 93, able to view ad visuals like a person and scan an entire category's creatives in ten minutes.

A long post collects Astra's uses in modeling, indie games, and video editing, aimed at people who don't want to burn tokens on trial and error.

A quick reaction from someone who now has Astra access through ChatGPT Work and Codex.

A user asked Astra to improve a landing page's Lighthouse score, but it renamed a docs nav label and added an unrequested dark theme; planning feels more nuanced, but it means more code to review.

A video shows a GPT-6 Astra prompt saying the usage limit has been reached and will reset at 7 AM, reflecting tight limits early after launch.

A link to OpenAI's official post, which lays out how GPT-6 Astra is offered across ChatGPT Work, Codex and the API.

A leak says OpenAI is internally testing GPT-6 Sol, which is fast but below Astra, may launch at DevDay on September 29, and could be a performance-and-limits middle ground for Plus users.

A user asks Grok directly, in front of OpenAI, for its most honest opinion of GPT-6 Astra.

On a new electrical engineering benchmark, Astra takes first place at $0.83 per task, 2 minutes end to end and 6k tokens, while Grok 4.6 performs surprisingly well.

A Japanese piece covers OpenAI's GPT-6 Astra release and focuses on ARC-AGI-3, where Astra scores 99.9% with external tools but only in the 60s without them.

A tester says GPT-6 Sol's output is clearly worse than Astra overall but fast, and it produced the result zero-shot in 15 minutes and 60k tokens at Max effort.