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

In a video, Google DeepMind's Tulsee Doshi says Argon's 1M-token window is not meant for writing books but for deep reasoning tasks like scientific research and open math problems, which a 64K cap would cut short.

A writer who previously found Gemini the most monotonous (13 spots where the same sentence ending repeated three times in 17 sentences, versus 1 for ChatGPT and 9 for Claude) plans to rerun the same test on Argon today.

The author stresses this is not a generated video but a .blend 3D scene file produced directly by Gemini 4 Argon through Blender MCP.

A leak claims that while training Gemini 4, codenamed Argon, Google accidentally fed in chain-of-thought data from its own models.

DeepMind's Gemini 4 Argon infers spatial and full-body motion from a video of tossing a bag into a hole, auto-generates a physics simulation, and the humanoid robot lands 7 throws in a row.

A user says treating Argon the way they treat Claude and GPT nearly drove them to a breakdown, and argues Argon is likely stronger than shown, with the problem lying in user habits rather than the model.

A user shares a long, A+ grade report Argon generated on the initial reaction to its own release, calling its real-time news digestion extremely strong.

The Google safety researcher says the Argon safety report is Google's most thorough by far and recommends its loss of control, alignment and monitorability assessments.

A user shows Gemini 4 Argon's instance segmentation results, saying it correctly produces masks that include holes and disconnected parts of the same object.

After getting access, the author had Gemini 4 Argon turn their camera roll of cursed cat memes into Kitty Combat, a fighting game with 12 cat characters and a CATALITY finisher, currently only a gameplay trailer.

A news roundup says Gemini 4 Argon is available to paid API tiers and Pro/Ultra users, ranks 5th overall on Arena and sets a new SimpleQA Verified SOTA, though self-testing puts it behind Fable and Astra, and Cursor pricing doubles on 12/1.

A user says Gemini 4 Argon made its own promo video in HyperFrames and performed well on their Code2Video Bench.