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

A developer says they built InvoiceGen in two hours using Gemini 4 Argon via API, found it clearly better than Gemini 3.8 Flash, and asked when it will come to Antigravity.

A short comment argues that Gemini's weak harness layer wastes the capability of the Argon model itself.

A user posts images saying Gemini 4 Argon "definitely has vibes," showing the model's actual output.

A general remark that people are slowly realizing how good Gemini 4 Argon is, with no specific benchmark data.

A post showing Gemini Argon explaining how it figured out how to play cornhole, which the poster calls super cool.

A user who used a full 5-hour Gemini Pro window found it consumed only 5% of the weekly quota, meaning 20 sessions a week or about 14.3 hours a day to exhaust it.

A post says Gemini 4 Argon launched October 13 for complex multi-step tasks, scoring 77.9% on the DeepSWE coding benchmark ahead of GPT-6 Astra and Claude Opus 5.5, priced at $2/$10 per million tokens before doubling after November 30.

A short post announces that Gemini 4 Argon has finally been released and asks for real-world coding comparisons against Opus 5.5.

Google's official blog introduces Gemini 4 Argon, claiming a faster quantum error decoder and an arcsin circuit that is 100x cheaper.

RouterLink promotes Argon as the first Gemini 4 model, aimed at long-horizon coding, finance, legal and science work, with a 1M-token context window and up to 1M tokens of long-decode output, and says it ranks first on the Vals Index.

Someone built a Subway Surfers-style playable game with Gemini 4 Argon through Zevio Harness, which claims inference costs up to 45% lower than official pricing.

Someone used Cursor with Gemini 4 Argon to build a multiplayer 3D capture-the-flag game with cartoon animal characters, which the author sees as Google's strong return to game development.