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

A developer used Devin with Gemini 4 Argon to generate a Minecraft clone in 19 minutes, using 6.3M tokens and about $3, and says the model looks strong at design and 3D tasks.

A blogger posts a video calling Gemini 4 Argon a "bummer" and uses hands-on footage to show dissatisfaction with the new model.

In AI Studio, a single prompt to Gemini 4 Argon produced an interactive physics intro page with electromagnetic field, gravity funnel animations and sliders, taking 3 minutes 22 seconds.

Given the same prompt to write code that generates a MIDI seaside-stroll song, Gemini 4 Argon beats both GPT versions but falls clearly short of Opus 5.5.

A user says Gemini 4 Argon is directly available in Devin, tested it with two SVG animation prompts, and calls GPT-6 Astra and GPT-6.1 Sol clear losers, with video results attached.

Third-party platform Pollo AI announced that Gemini 4 Argon is now available on Pollo Agent, adding another option to its model roster.

The user calls Gemini 4 Argon a successful SOTA-level model that now sits at the same table as Opus 5.5, and takes a jab at OpenAI along the way.

Developer Ammaar says Gemini 4 Argon can generate 3D assets and game mechanics from scratch in one shot, and shared prompts and code for six playable, remixable games.

On the same 3D task, Argon finished in 15 minutes for about $5 while Opus 5.5 took 30 minutes and about $12.50, and the author found Argon's result more colorful and detailed, nearly 4x faster and less than half the cost.

Hands-on tests say Argon ranks first for object detection in dense scenes and for extracting small, handwritten, rotated, and table text, at a quarter of Astra's cost under high load.

saloniiio says they waited a long time for Gemini 4 Argon only to find they cannot afford to use it.

LyalinDotCom says Gemini 4, codenamed Argon, is coming out today, answering earlier questions, though no official source is given.