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

A user argues Astra is a research model aimed at 5% of users, so what matters for Gemini 4 is Google's architecture and pricing, and hints an update will come soon.

A Japanese post says Gemini 4 Argon benchmarks are competitive with GPT-6 Astra and Fable 5.1, but suggests Google may have limited performance after public release due to compute constraints.

A commentary says Gemini 4 beats all rivals including Fable and Astra on key benchmarks, asking whether Google is back at the frontier, though only on paper so far.

A Spanish-language post says Google announced Gemini 4 Argon today as a frontier model rather than a Flash model, with 1M output tokens and strengths in long coding, legal, finance and cybersecurity, currently limited to testers.

A benchmark roundup puts Argon at 77.9% on DeepSWE, tied for first at 68% on CWE-bench and first at 51.3% on AutomationBench, with its output limit raised from 64K to 1M tokens and access starting with the Fairwind program and Ultra subscribers.

Contrasting 2023 with 2026, the author says Gemini 4 beats GPT-6 and Claude on nearly every benchmark and is bullish on $GOOGL.

A report says Gemini 4 scores well on benchmarks but holds up less well when Google employees use it in practice, especially on certain coding tasks.

Google DeepMind released Gemini 4 Argon, the first Gemini 4 model, raising output from 64K to 1M tokens and targeting long-horizon coding, enterprise knowledge work and cyber defense, though it trails Astra and Opus 5.5 on some benchmarks.

On AA-Omniscience hallucination rates, Gemini 4 Argon scores 15% versus 51% for GPT-6 Astra and 66% for Opus 5.5, with the poster arguing that knowing when to stay quiet is the ultimate benchmark.

midudev summarizes Google's Gemini 4 Argon launch, placing it on par with GPT-6 Astra and Fable 5.1 and noting it is currently available by invitation only.

Arena's official account says Gemini 4 Argon (High) reshapes the Agent Arena Pareto frontier at $0.62 per task, ranking #8 with a +7.92% net improvement score.

A leak claims Gemini 4 reaches 45.3% on PostTrainBench, more than double Gemini 3.1 Pro's 21.99% and above GPT-6 Astra.