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

A blogger notes DeepSeek V4.1 Flash-1017, a lightweight model, is being pitted against Opus 5.5, Gemini 4 Argon and GPT-6 Astra.

DeepMind's Oct 13 video says Gemini 4 Argon watched cornhole footage, built a physics sim and drove Apptronik's Apollo 2 to seven straight shots, but no attempt count is given, so the author calls it a highlight not a rate.

The author gave DeepSeek V4.1 Flash and Gemini 4 Argon the same one-shot task: Argon took about 10 minutes and 2% of weekly quota, V4.1 Flash about 4 minutes and free, but he still rates it below Argon.

On a coding-agent benchmark spanning 11 unfamiliar ML and data-analysis tasks, Gemini 4 Argon places fifth at 0.202 while GPT-6 Astra leads at 0.422.

A Persian video shows using Gemini 4 Argon with the Canvas tool to convert an uploaded PDF into an animated, interactive web app via a prompt.

Gemini 4 Argon (high) scored 20.2% on WeirdML, landing between Sonnet 5.5 and Fable 5.1, strong but still behind the frontier on token efficiency.

A developer advises against using Gemini 4 Argon to write complex systems directly, praising it for building tools and reviewing Opus's tech choices, suggesting a 300k auto-compaction cap and noting it clashes with GPT.

A short poll asks readers to pick the best coding model right now among GPT-6 Astra, Opus 5.5 and Gemini 4 Argon, and explain why.

A Japanese developer tested how far Claude Code can build a video using only code, and had Gemini 4 Argon write the prompt itself; the result video is attached.

In a same-prompt test where six models designed 3D-printed capsules to protect a raw egg, Gemini 4 Argon placed third with a slight crack at 4 ft; only Opus 5.5 and GPT-6 Astra cleared 4 ft.

A user reports Gemini 4 Argon made limited progress and many mistakes building on a compiler, needing Fable to clean up, while Meta Muse Wonder Full made none in a similar test.

A short post says Gemini 4 Argon's main remaining problem is its quota, which can't support intensive work.