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

Google put out benchmark numbers directly: Finance Agent 65.4% vs 53.5%, Legal Agent 19.6% vs 5.4%, Vibe Code 91.9% and GraphWalks 99.7%, but it still trails on SWE, Terminal, Science and OSWorld.

Someone spotted Google's gemini-4-argon in Artificial Analysis's new model listings, an unannounced leak signal.

A chart comparing cost efficiency across three flagship models puts Gemini 4 Argon right between GPT 6.1 Sol and GPT-6 Astra.

A commenter argues that current employees leaking to Bloomberg on launch day that Gemini 4 is weak at coding shows how dysfunctional Google's internal culture has become.

The post says Gemini 4 Argon leads Claude Opus 5.5 and GPT-6 Astra on many tasks, opens first to cyber defenders via the Fairwind Program, and will roll out to developers and enterprises.

After trying Gemini 4 Pro on Arena, the author says its explanations feel highly pedagogical, interactive and fast, but whether it can complete tasks as efficiently as the Sol models remains the key question.

A Thai long post rounds up Argon's scores on DeepSWE, AutomationBench, Vals Finance and Harvey's legal benchmark, plus internal uses like 1M-token output, a Rust migration in the Fuchsia kernel and automated cyber vulnerability patching.

Google says Argon matches or beats OpenAI and Anthropic models on coding and cybersecurity benchmarks, and is opening it first to selected companies and governments through its Fairwind programme, with no public release date.

A post says Argon is Google's answer to GPT-6 Astra and Claude Fable 5.1, leads on almost all benchmarks, and will roll out first to Google AI Ultra subscribers and via API.

A post says Gemini 4 Argon has reached the #1 position on the Text Arena leaderboard, with a screenshot of the ranking.

A post says Gemini 4 Argon will go from internal use to trusted testers and then Ultra and the API, with 1M output tokens, strength on long-horizon coding and agent workflows, and a claimed 77.9% on DeepSWE.

Charts show Argon ahead with 77.9% on DeepSWE, 91.9% on Vibe Code Bench, 88.8% on LABBench 2 and 84.2% on GraphWalks, while Astra still leads FrontierSWE and others.