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

A demo video says Argon built 3D spatial coordinates and simulation code from raw video of a human playing cornhole, then sank seven consecutive shots after transfer to a humanoid robot.

A developer says Gemini 4 Argon is now available worldwide in AI Studio, the Gemini API and Antigravity, and calls it state-of-the-art after weeks of building and a test at PyCon Africa.

A developer says Argon responds quickly enough for everyday use and jokes about how long their token budget will last.

The third-party gateway Merge Gateway has added Gemini 4 Argon, which scores 84.2% on GraphWalks at 256K to 1M tokens, 12 points ahead of the next model, and 77.9% on DeepSWE v1.1.

An analysis argues that Gemini 4 Argon's low running costs push buyers to weigh migration work against savings, so switching costs could decide who benefits from cheaper AI.

Gemini 4 Argon is live today across the API, Gemini app, Search, Antigravity and Enterprise, ranking first on the Vals Index with 1M context and $2/$10 intro pricing.

Melius MCP announced support for Gemini 4 Argon, citing 51.3% on AutomationBench and 91.7% on LVBench for long-context multimodal comprehension.

A user posted a video claiming Gemini 4 Argon generated a game in just 19 minutes, saying game developers would be speechless.

A user shared screenshots saying Gemini has the strongest emotional understanding and that they are very happy with Argon so far.

A user says Gemini 4 Argon performs well on world understanding tasks and unlocked a new benchmark test.

A short Japanese post with an image states that Gemini 4 Argon's level of intelligence is now confirmed.

NovaXCode posted an image captioned "gemini is so back with argon 4," expressing excitement about Gemini 4 Argon without giving any evaluation details.