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

A user complains that frontier model access lasted only five minutes and hopes Google releases Gemini 4, codenamed Argon, quickly.

Google is reportedly A/B testing Gemini 4 in Antigravity under the codename Gemini-3.8-flash-high, with Ultra and API users getting it first, plus a generated pelican-on-bicycle video.

In an All-In interview, Sundar Pichai praised Elon Musk's ability to bring future tech into reality, read as setting the bar for Gemini 4 Argon's push into complex workflows, cyber defense and software engineering.

A rumor says Tibo teased GPT-6.1 Astra, Fable 5.5 is Anthropic's response, and Argon shipped September 30, giving Google an early lead this round.

The streamer says a 48-hour stream will be followed this week by a tournament mixing Terran, Zerg and Protoss matches, and Gemini 4 Argon might play.

A Japanese post says Google released Gemini 4 Argon with a 1 million token output limit per run, about 8x that of GPT-6 Astra and Claude Opus 5.5.

The author uploaded a video as input and says Gemini 4 Pro generated this new video from it, showing its video understanding and generation.

Jonah Blake says Google's Gemini 4 Argon scores better than Fable and Astra on almost every test, and notes Google Ultra costs him over $200 a month with no return.

A leaderboard screenshot shows Gemini 4 Argon performing well on a coding ranking, and the poster says "Google is back."

An Argon-generated overhead shot of a Mars rover, with tire tracks, dusk fog lights and terrain detail, is called one of Argon's best single outputs so far.

Another video labeled as generated by Gemini 4 Pro, a short creative clip titled "Cat," offered as a reference for the model's output quality.

A user demonstrates motion design made with Gemini 4 Pro in an attached video, showing its animation and visual generation results.