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

Google released Gemini 4 Argon, claiming #1 on the Vals Index, internal use for refactoring 6 million lines of Chrome code and nearly 3x faster Gemma, priced at $2/$10 against Sonnet 5.5.

A leak says Gemini 4 Argon beats GPT-6 Astra, Claude Fable 5.1 and Opus 5.5 on most listed benchmarks, including 77.9% on DeepSWE v1.1 and 99.7% on GraphWalks.

On the same prompt at the highest reasoning setting, Argon and Opus 5.5 gave nearly identical results, with Argon taking 13 minutes and $1.2 versus Opus at 15 minutes and $8.2.

A user jokes that Google released Gemini 4 Argon on Antigravity without letting users actually use it.

A first-impressions post says Argon needs some handholding but is state of the art in design and frontend, and the author bought Google Ultra to try it and may pair it with Opus 5.5.

A user points out that Gemini 4 Argon is not available to Pro subscribers on Antigravity and is limited to higher tiers.

The author had Argon edit raw assembly code to build a donut shop level in an old Super Nintendo game, complete with new pixel art, double jumps and a looping ragtime song.

A Persian-speaking user's first test of Gemini 4 Argon finds chat output highly detailed but slow, with pricing at about half of Opus 5.5 and real quality still to be proven in coding.

Firebase says developers on the paid tier of the Gemini API can call Gemini 4 Argon client-side across Android, iOS, Flutter, Unity and web apps via Firebase AI Logic.

A user shared a video of a complex 3D scene generated by Gemini 4 Argon, calling it the best Argon output they have seen, and said they want to try a more complex Indian temple model but their laptop cannot render it.

TestingCatalog shared the Arena overall leaderboard, where Gemini 4 Argon ranks fifth behind Opus 5.5, GPT 6.1 Sol, Fable 5.1 and GPT 6 Astra.

A Japanese user announced the arrival of Gemini 4 Argon with an image that appears to be a screenshot of the interface, offering a first look.