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

A Chinese user says Gemini 4 Argon's copywriting and frontend work is decent, but Pro hits limits quickly and struggles with complex problems, not enough to justify upgrading to Ultra.

ARC Prize released Gemini 4 Argon's full ARC-AGI results, with links to the leaderboard, a reproduction repo and its testing policy.

Developer markessien says pairing Claude with Gemini 4 Argon works best, better than Claude with GPT, without giving test details.

A user spotted traces of Gemini 4 Argon in the Antigravity CLI and guesses that Pro subscribers will get access before long.

On the Gemini app's Pro plan, Gemini 4 Argon built a complete Three.js runner game in one go, with 3D scenery, trains, buildings, animations and audio.

A Medium article argues against defaulting to the Medium thinking level and says effort budgets should be tuned separately for Gemini 4 Argon and 3.8 Flash.

On the same prompt, Muse took about 40 minutes and cost around $30, while Gemini 4 Argon took about 30 minutes and cost around $10, and the author thinks Argon did better.

Google says Isomorphic Labs researchers use Gemini 4 Argon to parse complex PDFs and patents, cutting information extraction from weeks to hours.

A user who tried Gemini 4 Argon says the experience is much better than GPT-4o, without giving specific test details.

The author had Gemini 4 Argon, GPT-6.1 Sol Max and Sonnet 5.5 each build a voxel Porsche 911 GT3 RS in Three.js, found Claude's free model did better, and sees no reason to leave Claude and Codex for Gemini 4 yet.

A roundup of 8 unexpected Gemini 4 Argon uses, including voice generation from raw Python, video-to-SVG conversion, and a game about something huge arriving.

A video demo shows Gemini 4 Argon looping through code-writing retries (425,250 tokens in one attempt) while Claude Opus 5.5 reviews and signs off, with the poster noting Argon still returns 404 for most people.