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

Gemini 4 Carbon launched two weeks after Argon and Google says it beats Opus 5.5 on most tests, but access is limited to a small group of cyber defenders with no date for everyone else.

A user says Gemini 4 Carbon is out but its benchmarks are as disappointing as Argon, an unverified negative take.

A user posts Gemini 4 Carbon's official benchmark results and claims it does what OpenAI couldn't, edging out Opus 5.5.

The post says Gemini 4 Carbon is Google's new SOTA model with 65.2% on Terminal-Bench 4.0 and $2 input/$10 output per 1M tokens, but only a small group of security users gets access first.

The user notes months of waiting for Argon, then Google shipping Gemini 4 Carbon only two weeks later, marking a sudden acceleration in releases.

Carbon leads AA at 59 vs Opus 5.5's 58 and GPT-6 Astra's 57, scoring 88.2% on vulnerability discovery, 54.8% on exploitation and 71% on patching, with access starting for select cyber defenders.

A user says Argon caught their attention but the rollout was badly handled, leaving them mentally writing off Google's models.

A Spanish-language roundup says Carbon leads on AA index, Terminal-Bench 4.0, long context and cybersecurity, loses only OSWorld to Opus 5.5 and FrontierSWE to Astra, and is limited to Fairwind cyber defenders with no public pricing yet.

A user reports Gemini 4 Carbon is out with better performance than Argon and is currently rolling out to cyber defenders.

The poster says Carbon outperforms GPT-6 Astra, Claude Opus 5.5 and Fable 5.1, far exceeding the earlier Gemini 4 Argon.

A roundup says Gemini 4 Carbon beats GPT-6 Astra, Claude Opus 5.5 and Gemini 4 Argon across general intelligence, agentic coding, terminal, long-context and security benchmarks, citing 96.2% on GraphWalks and 65.2% on Terminal-Bench.

A Chinese user says Gemini 4 Carbon tops Opus 5.5 on almost every benchmark, though access is currently limited.