Gemini 4 Pro / Flash: benchmarks, hands-on tests and leaks

On the same prompt with /goal, Gemini 4 Argon took 38 minutes, far faster than Fable 5.1 at 2h 10m and Astra at 1h 25m.

A post announces that Gemini 4 Argon is out and jokes that the first "Argon is cooked" take from someone who used it for seven minutes will follow.

GitHub's changelog confirms Gemini 4 Argon is available in Copilot for Pro+, Max, Business and Enterprise users across major IDEs and Copilot CLI.

A short video shows Gemini Argon carrying out debugging tasks on a robot, presented as a hands-on demo.

vikktorrrre posts a video claiming Gemini 4 Argon outperforms every competing model at 3D work and that Google has made its comeback.

A video compares 3D modeling output from Gemini Argon 4 and Sonnet 5.5, and the author gives the edge to Argon 4.

A Google developer account shows Gemini 4 Argon orchestrating subagents and using Nano Banana to generate assets, turning a desk photo into an interactive 3D racing game.

The author says Gemini 4 Argon trades blows with GPT-6 Astra with almost no visible difference in output quality, but costs 5x less, and promises hands-on tests soon.

The author had Gemini 4 Argon generate an intro video from scratch in one shot for $4, calling it capable but with odd design sense.

Cline says Argon ranks first on the Val Index for agentic tasks at half the cost of Opus 5.5, and Google says the Chrome team used it to refactor about 6 million lines of legacy features out of a 40-million-line codebase.

A user's test with the same prompt found Gemini 4 Argon generated faster than every model it was compared against, with a video demo attached.

A demo shows a robot learning from its errors and gradually working out how to complete a task, which the poster calls the kind of Gemini 4 Argon demo they want to see more of.