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

The author tested Gemini 4 Argon on a self-built Noahbench and says it is nearly useless on real-world code, speculating that Google has given up the coding model race to target knowledge workers instead.

A preview post announcing the author will join Theo Jaffee's livestream to discuss all things Gemini 4 Argon.

An Arabic-language post shows Gemini 4 Argon taking a Pixel CAD file and using Blender MCP to produce an assembly animation, lighting, camera moves and audio, with 1,000 iterative verification passes.

A post says the video was made with Gemini 4 Argon and remarks that this has been the worst year for film and VFX studios.

Google has opened Gemini 4 Argon from closed testing to all AI Pro subscribers, citing 68.9% on Vals Index and 77.9% on DeepSWE, with promotional API pricing of $2/$10 per million tokens.

A Brazilian user upgraded to Google's Ultra plan to test Gemini 4 Argon and stream the tests live, saying the only feeling so far is regret over the money.

A purely emotional announcement that Gemini 4 Argon has launched, with no details, screenshots or test results.

Google's lmthang shared a screenshot showing Gemini 4 Argon at #5 on the Agent Arena leaderboard, saying it is still climbing.

After testing, a user says Gemini 4 Argon is good with visuals and complex documents but still lags Opus 5.5 on coding and agent tasks, and feels less consistent overall than GPT-6 Astra.

Max Bittker says Gemini 4 Argon performs similarly to Astra on the Runescape Bench but costs only as much as the Sol tier.

Phil Schmid gave Gemini 4 Argon a photo of the Manhattan skyline, and it wrote a multi-pass canvas program that builds the oil painting from broad background strokes to fine water highlights.

Google says Gemini 4 Argon scored 68.90% on a finance, coding, legal and tax benchmark, ahead of Claude Opus 5.5 and GPT-6 Astra, at $2 per million input and $10 per million output tokens, doubling after November 30.