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

The author speculates Gemini 4 Flash will be extremely fast, stay cheap, focus on computer use and agentic tasks, or be scrapped in favor of another Argon version.

A reply ranks computer-use price-performance as Kimi K3.1 > Gemini 4 Argon > Muse Wonder > Sol 6.1.

A developer demos MicroDuck, a home robot planned with Gemini 4 Argon and trained on DGX Spark, detecting falls without sending camera frames to the cloud.

A user says Gemini 4 Argon in Antigravity used /browser to open Chrome, download forms, merge scans into a 17-page PDF and file a full hospital insurance claim.

A developer found computer-use music production in Logic Pro too expensive versus SUNO, and after Gemini 2.5's audio fell short plans to test the newer Argon.

Hyperbrowser now supports Gemini 4 Argon, letting it browse, click and type autonomously to complete multi-step tasks in Gemini Computer Use.

A new independent benchmark tests strategy and computer use in Stronghold Crusader, with Gemini 4 Argon at 62.3 ahead of GPT-6.1 Sol's 56.9, and all code and recordings released openly.

Roboflow says Gemini 4 Argon expands what is possible in computer vision and stresses routing tasks to frontier or specialized models by accuracy, speed and cost.

Argon reportedly watched a single human cornhole video, wrote its own physics simulator, trained a throwing policy and transferred it to the Apollo humanoid for 7 hits in a row, cutting weeks of robotics work to hours.

A Russian-language post claims Gemini 4 Argon is fully multimodal, sees, hears and speaks, and after simulation training is ready to deploy on humanoid robots.

Cua says it had early access to Gemini 4 Argon, citing major gains in multimodal reading of screens, documents and spreadsheets, and has integrated it into Cua Driver to act inside apps.

Third-party platform AIHubMix has added Gemini 4 Argon, advertising a 1M-token context, multi-step planning, Active Memory and Active Focus for coding and long-running agents.