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

A short observation that Gemini 4 Argon has made real progress on 3D-related tasks.

A roundup of evaluators' and developers' checks on Gemini 4 Argon finds it roughly matches GPT-6 Astra at about 60% of the cost, though some say it is not first and is weak at editing code and self-checking.

Argon (high) scores 98.6 on the Extended NYT Connections Benchmark, up from Gemini 3.1 Pro's 97.4, with cost per puzzle down 71%.

A collection post that claims to gather all available information on Gemini 4 Argon for anyone who wants to catch up in one read.

Google AI shared an experiment using Gemini 4 Argon to generate training simulations from human action videos, showing a robot learning a beanbag-toss task, with the author urging questions about where the simulation's physics parameters come from and when they fail.

A Japanese user who gave Gemini 4 Argon a design task requiring web research says it still carries its old habit of assuming a cause up front and investigating sloppily.

A Japanese user shows motion graphics made with Gemini 4 Argon in the Gemini app's Canvas, complete with narration generated by the free Gemini TTS (2.5 Flash) inside Canvas.

Argon (high) scores 81.2 on BuildingBench, up from Gemini 3.8 Flash's 54.8 and sixth among 30 models, at roughly one-fifth the cost of Fable 5.1.

A user ran the same prompt through Gemini 4 Argon and Opus 5.5 to generate 3D content, posted a video, and called the winner obvious.

ValsAI asked Chinese and American models hundreds of politically sensitive questions about China and said Gemini 4 Argon, Mistral Large 4 and Grok 4.7 occasionally repeated official propaganda lines.

A user generated a "pelican riding a bike" image with Gemini 4 Argon on mobile and shared the result: the pelican had no legs.

A review says Gemini 4 Argon completes tasks very efficiently and thinks things through thoroughly, but its visual output is not impressive, feeling weaker than Anthropic's models and stronger than OpenAI's.