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

A video puts Gemini 4 Argon and Claude Opus 5.5 side by side for anyone who wants to see the two frontier models go head to head.

To mark Argon's launch, the artist drew a new spirit-armor illustration of the original character Miraluna, themed on deep thinking and connecting to the future.

A video runs the same rocket launch task through Gemini Flash 3.8 and Gemini 4 Argon to see which model does it better.

A user asked Gemini 4 Argon to generate a town, called the result shocking, and joked that Claude is done.

A developer says the benchmark isn't finished yet, but early results show Gemini 4 outperforming Opus 5.5 and GPT-6 Astra on agentic tasks in a Turkish legal context.

A tester says Argon's coding is adequate for general needs but clearly slower than Opus 5.5, and that the antigravity app has no built-in browser.

The weekly Antigravity SDK update adds gemini-4-argon model targets (still defaulting to gemini-3.8-flash when unspecified), multi-agent workflows, in-memory skill definitions and output limits beyond 64k tokens.

A Japanese user used Gemini 4 Argon (medium) to generate a local city information website in one pass, saying the quality is good enough to launch as is.

In a one-day test, Gemini 4 Argon tied Gemini 3 on six everyday questions, ran 2-6x slower and invented a source, but its task mode researched on its own for 7 minutes and delivered a 15-row comparison table with numbers that checked out.

A user had Gemini 4 Argon redesign the withrelic.com dashboard and called the result insane, sharing a video of it.

A user gave Gemini 4 Argon a single prompt and 180k tokens to produce a 49-minute space travel video, then asked people to rate it from 1 to 10.

A user asked Gemini 4 Argon to build a realistic Earth with the moon, landmarks and the ISS in Blender; it took about 1 hour 32 minutes and cost roughly $90, with 3D work rated on par with Astra but below Opus.