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

On the new RoboQuest robot agent benchmark, Gemini 4 Argon scores 23.8% success at $11.22 per episode, roughly level with GPT-6 Astra and well above Opus 5.5's 13.8%.

A blogger lists three reasons Kimi's next model can't afford to lose, noting that Opus 5.5 and Gemini 4 have pushed both capability and pricing to the limit.

A joke post complains that quota resets aren't enough to satisfy Sad Albert, and says DeepMind's Logan failed to do so with Gemini Argon.

A Persian post says Google asked Gemini 4 Argon to build a 3D library in Python with natural lighting and realistic detail, and the model produced about 5,000 lines of code, implementing even the lighting system and a video tool itself.

A developer asked Gemini 4 Argon to build an animated Excalidraw clone from scratch, says the result is fully working, and plans to open-source it.

A Polish developer says that in audits of T3 Code and a large Rust codebase, Gemini 4 Argon made false claims about how the systems work more than half the time, and also changed frontend code and file structure that were not requested.

The author says a single prompt to Gemini 4 Argon produced an entire launch video for HyperFrames.

Developer markessien says Gemini 4 Argon runs "very slow," contrasting with his same-day praise of it paired with Claude.

A user who says they know nothing about coding claims Gemini 4 Argon helped them build 3 working web apps in 30 minutes, posting local port addresses.

A user asks whether Argon's 1 million token context lets AI reason reliably across a massive codebase, or just gives it more information to miss.

A user says Gemini 4 Argon produced an entire music video in one go and also wrote the prompts used for the music.

A joking post imagining how Gemini Argon would react when working on code generated by Meta's Muse Wonder.