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

A Vietnamese thread says Google released Gemini 4 Argon early on October 1 for long coding, finance, legal and cybersecurity tasks, limited to trusted security organizations, topping or tying 13 of 18 self-reported benchmarks at $2/M input and $10/M output.

A screenshot shows Gemini 4 Argon at #1 on Text Arena with Opus 5.5 pushed to second, priced at $2/M input and $10/M output, potentially making it one of the cheapest frontier models.

A user who has been critical of Google models says Gemini 4 looks strong on benchmarks and hopes Google pairs it with a first-class harness like Codex.

DeepMind released its flagship Argon model, raising the single-response limit from 64,000 to 1 million tokens and scoring 77.9% on DeepSWE v1.1, aimed at long-horizon software engineering and enterprise research.

A developer says Gemini 3.8 Flash was underrated and was used to build a hyper kernel foundation for Strix, and plans to return to it once Argon (Gemini 4) access is granted.

smhn reports that Google released Gemini 4 Argon, claiming performance beyond Astra and Fable, with early access given to cybersecurity defense organizations.

A poster argues Argon has not really launched, is limited to enterprise and Ultra subscribers, and that by the time Pro users get it, Astra and Fable will likely have moved two iterations ahead.

Google says Argon leads or ties Astra and Opus on 13 of 18 public benchmarks, but the first wave goes only to Fairwind partners with cyber guardrails turned off.

Google has reportedly released Gemini 4 Argon, scoring 19.6% on legal evaluations versus 5.4% for Astra and 3.8% for Opus 5.5, currently limited to cyber defense partners with no public release date.

A post says Argon can autonomously detect and patch security vulnerabilities, beats GPT-6 Astra and Anthropic Fable in benchmarks, and already saves hundreds of terabytes of memory in Google's own data centers.

A Japanese user says Gemini 3.8 Flash's Japanese is natural enough to need no humanizing, looks forward to Gemini 4 Argon, and wonders why it isn't named Gemini 4 Pro.

A post claims Google skipped Gemini 4 Pro and released Gemini 4 Argon directly, raising output to 1 million tokens and pricing it at $2 input / $10 output, about 80% cheaper than GPT-6 Astra.