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

PromptLLM shares a screenshot of what it calls an insane take from Gemini 4 Argon, without stating the content.

A Gemini 4 Argon agent in AI Studio researched, iterated and added Lyria music on its own, producing a game 20,000 people have played.

After 329 sessions, 5.86 billion tokens and 404 hours of autonomous work, the author read every log to write an honest Gemini 4 Argon review covering the good, the odd and what still needs fixing.

The author suspects Minimax M3.1 Flash was trained on AA Omniscience questions after it outscored Gemini 4 Argon, and a paraphrased question test shows it reciting the original answer.

A community note on Antigravity's Gemini 4 Argon tweet reflects user worry that Google may split Pro subscribers into separate tiers.

Citing The Information, the post says Fable 5.5, Gemini 4 Carbon and GPT-6.1 Terra won't arrive this month and are expected in November.

The post says Carbon is an upgrade over Argon arriving this month, notes Google's strength in visual tasks, and hopes it can genuinely compete with Astra and Fable rather than being incremental.

Every reports that Dan Shipper had his agent ask Gemini 4 Argon to propose structural edits, then apply them as tracked changes in Google Docs, roughly doubling his writing speed.

The Information reports Anthropic is preparing a model better at inferring unstated intent, and notes Google is working on Gemini 4 Carbon, an upgrade to Gemini 4 Argon.

After watching a human play cornhole, Gemini 4 Argon drove Apptronik's Apollo 2 to sink seven straight shots and generated 4,500+ sim episodes lifting success from 56% to 75%.

Someone had Gemini 4 Argon synthesize licensed CS2 and Skate 4 into one game, running Dust 2 layout on Skate 4 physics with no custom map or mod pack.

An AI Ultra subscriber reports that 1–2 hours of regular coding with Argon High in Antigravity left the weekly limit at 21% and the five-hour limit at 35%.