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

The Vals RSI Index measures whether a model can take part in developing the next generation of models; Argon rose from last place to 4th, ahead of GPT-6 Astra, though the top score is only 37%.

A Google DeepMind staffer says Gemini 4 Argon is the next-generation model built for cyber defense, leading on prompt injection robustness in the Gray Swan IPI benchmark.

The author says Gemini 4 Argon opens to 650+ security partners through the Fairwind Program with cyber guardrails removed, after passing a voluntary US government pre-release review, following the same playbook as GPT-6 Astra and Claude Mythos.

A post says Gemini 4 is state of the art on knowledge benchmarks, with coding performance roughly on par with Astra/Fable 5.1, and worries that safety review could delay launch.

A Japanese post rounds up Google's published figures: performance close to GPT-6 Astra and Claude Opus 5.5, output limit raised from 64K to 1M tokens, and pricing of $2 input and $10 output per million tokens.

A Japanese news flash says Google released Gemini 4 Argon, which beats GPT-6 Astra and Claude Opus 5.5 on some development, finance and business automation tests, raises the output limit from 64K to 1M tokens, and is first offered to cyber defense organizations.

A poster claims Gemini 4 outperforms GPT-6 across the board at 80% lower cost, and beats Astra, Fable and Opus 5.5 on coding.

On Artificial Analysis's cost-versus-intelligence chart, Gemini 4 Argon scores about 52 at $1.99 per question, versus $3.26 for GPT-6 Astra and $7.63 for Claude Fable 5.1, though GPT-6.1 Sol offers similar scores at $0.72.

A user shares three outputs from the Gemini 4 Argon checkpoint during its anonymous Arena testing, saying all three outperformed GPT-6 Astra and Claude Opus 5.5, and is waiting to test the official release.

Google officially announced Gemini 4 Argon to replace the entire Pro line, but the model is gated behind a closed cybersecurity initiative called the Fairwind Program, and the lack of a public release dropped a prediction market contract from 51% to 14%.

A tester says Gemini 4 Argon produces polished showcase pages in one shot but caps out quickly on ambitious tasks, and suspects it would not perform well on long-horizon benchmarks like METR.

A complaint post urging Google to release Gemini 4 Pro soon, warning that even Haiku alone could beat Gemini if the delay continues.