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

A post lists Astra's scores on ARC-AGI-3, FrontierMath, ExploitBench and OSWorld 2.0, and cites Greg Brockman's remark welcoming the AGI era.

A leaked benchmark sheet marked as predicted shows Gemini 4 leading GPT-6 Astra in math, terminal use, science and other areas, though Google has not published the numbers.

A leak says Gemini 4 began training in late July 2026, focused on agentic workflows and coding, and the poster puts its odds of beating Astra at launch at just 30%.

Astra is the first recent model to improve both workflow completion and restraint on Zapier's AutomationBench, leading on restraint at 68.3%, while Sol modified an extra frozen record.

A poster says ARC-AGI-3 went from about 0.5% to 99.9% in under six months and credits GPT-6 Astra, sharing a screenshot of the score.

An article traces desktop agents from MIT's 1979 Put-That-There to Astra's 72.6% OSWorld score, noting that ten chained steps succeed only 4% of the time at high cost.

OpenAI announced GPT-6 Astra, a new model that it says can control almost everything through voice commands, with a demo video.

GPT-6 Astra and 5.6 Pro have their own chat limits separate from Codex, so when Codex runs out you can connect those models to the Codex Harness through web quota.

Citing Astra's 57.9% versus Fable 5.1's 55.8% on Terminal-Bench 4.0, the author ran a scripted 20v20 agent battle in which the Astra side won 1-0.

A Japanese long-form piece reviews Brockman's September 3 remarks and the debate over defining AGI, noting that OpenAI's official documents do not declare AGI achieved.

A viewer jokes that OpenAI's Astra demo only shows someone ordering food and someone booking tennis courts, so most jobs are safe for now unless you're an executive assistant.

A September 4 roundup covering Astra's daily reset credit for paid users after access delays, Altman's apology, and praise for its coding value, 3D modeling, and PC control.