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

A user says Gemini 4 Argon took 4 minutes to answer a simple question, while ChatGPT answered it in 3 seconds.

DeepMind says four Gemini 4 Argon agents in a sandbox, using only code and text, wrote 2,783 programs, split work over chat and beat Minecraft in 27 hours with no human help.

A MineBench comparison pits Gemini 4 Argon against GPT-6 Astra, GPT-6.1 Sol and Claude Opus 5.5 on accurately recreating the Twitter/X feed.

A user generated a 2D image with Image 2.5 and fed it into the Argon Pro web version to produce a walking, wireframe-viewable, exploded-view 3D model of a wooden ox, saying its structure rivals Opus 5.5.

Argon was the best model on PPL's Wide-And-Deep-Research agent benchmark, with a lower task cost than Astra and Fable but higher than Opus 5.5.

A developer built a Skill in Antigravity on Gemini 4 Argon that lifts 2D isometric drawings into 3D mechanical animations and auto-generates explainer videos about the Three Kingdoms wooden ox.

The poster reminds readers not to mistake a certain piece of content for Gemini 4 Argon, noting that Argon rumors are being applied to things everywhere.

A post notes that users on the paid Pro plan can try Gemini 4 Argon now.

Marketing tool MigmaAi says it has integrated Gemini 4 Argon to auto-generate launch emails, paired with Mailcheap to control subscriber costs.

The author breaks down weekly Code quotas and regional shared-plan prices for Google AI Ultra tiers, warning that an unrestricted six-person plan can exhaust a week's quota in 3-4 days.

MTSlive recounts that four Gemini 4 Argon agents finished Minecraft end-to-end in 27 hours and 14 minutes with no human help, coordinating via in-game chat and recovering each other's items after deaths.

The user says pairing Gemini 4 Argon with Opus 5.5 produces stable output, but Argon's usage costs are high enough to threaten their Ultra subscription.