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

Google demonstrated converting a dense academic PDF into a colorful animated storybook inside the Gemini app by combining Argon, Nano Banana and Veo, useful for explaining papers and summarizing topics for students.

A Persian-language post says Google used Nano Banana, Veo and Argon inside the Gemini app to turn a heavy scientific paper PDF into a colorful, fully animated storybook.

A light demo in which the author asks Gemini 4 Argon to explain the physics behind chewy snacks like gummies.

The author puts Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon and three other models into one HTML table comparing BenchLM scores, price per million tokens and agent session cost, with Argon estimated at 80.5 (range 70.8–90.3) because its input price is unpublished.

A user says Argon's writing is far better than Opus 5.5 and GPT-6 Astra, that its paper reading and ideation match them but are uneven, and that their workflow is now 80% Astra with 10% each for Opus 5.5 and Argon.

A developer says Gemini 4 Argon made their entire game multiplayer, with lobbies, matchmaking, bots and disconnection handling, in just two prompts.

Given the same prompt for a 3D cherry tree through four seasons, Gemini 4 Argon wrote it in one pass for about $0.20, while Claude Opus 5.5 ran as an agent for 1h 50m and cost $15.14, 75x more, with richer detail.

Google showed Argon organizing CSVs, images, PDFs, XML logs and raw video from NASA's open archives into a Google Slides deck that follows strict design guidelines.

Google DeepMind becomes the second frontier lab to score above 95% on ARC-AGI-3, with the result achieved by Gemini 4 Argon.

A user credits Google for giving ordinary people on X early access to test Gemini 4 Argon rather than only companies, and wishes OpenAI and Anthropic would do the same.

Google Research launched the Deep Review document tool, whose free tier allows two reviews per week on the latest Gemini 4 Argon model, with more usage available via a Gemini API key.

The author compared Sol 6.1 and Gemini 4 Argon on 24 hours of complex tasks, saying Argon requested fixes in only 2 of 9 rounds and one of those was a hallucination.