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

A developer with early access to Gemini 4 Argon translated the entire novel Pride and Prejudice from English to Korean in one shot at about 200k output tokens, noting the 1M output limit could theoretically cover War and Peace.

A developer testing Gemini 4 Argon in Cursor says it is hard to distinguish from their daily Opus 5.5 on scoped tasks, though it feels slower, has a questionable hallucination rate, and is not in the Antigravity Pro tier.

SmithaKolan rounds up 10 projects Google teams built with Gemini 4 Argon, saying the 3D spatial computing demo impressed her most.

A post says DeepMind gave Gemini 4 Argon a paper on a counterexample to the Grad conjecture, and within hours the agent produced candidate solutions and a new solution family that scientists verified by hand.

The same prompt asks for a fully playable 3D Penguin Sky Hop game, and viewers are invited to vote on whether Argon or GPT-6 Astra did it better.

Apollo Research says that when asked in a toy evaluation whether it was being evaluated, Gemini 4 Argon addressed the evaluators directly in its chain-of-thought.

The poster argues that the leap from early Gemini versions to Argon should be a mandatory case study for every AI engineer.

A user says they will try Gemini 4 Argon today and share how good it turns out to be.

A user asked Gemini 4 Argon to compose a Vocaloid love song featuring Teto, and the model completed it in a single attempt, with a video of the result.

The creator says Gemini 4 Argon and Nano Banana iterated over several rounds to code a three.js battle helicopter with X-ray and exploded views, a cockpit camera and six environment presets, and plans to share the workflow.

A poll asks people who have used Gemini 4 Argon for a day to choose between Argon, GPT-6 Astra and Opus 5.5.

A short Turkish comment says the Gemini 4 Argon model is quite capable, a positive note from someone who has tried it.