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

A Japanese user celebrates Gemini 4 Carbon ranking first while complaining that the Argon and Carbon codenames are hard to tell apart and suggesting a rename.

An announcement-style post introduces Gemini 4 Carbon, citing strong gains over Argon from two weeks earlier in coding, cybersecurity and long-horizon tasks, now in trusted-partner testing.

Gemini 4 Carbon scores 59 on the index and 96 on long context versus Opus's 67, but still loses to Astra and Opus on FrontierSWE coding and isn't usable yet.

Jokes that Google names Gemini 4 models after noble gases, with Carbon following Argon and Krypton possibly next, while marveling at the absurd release pace.

Google released Gemini 4 Carbon two weeks after Argon; it leads the composite leaderboard at 59 and scores 96.2% on long context, but is limited to cybersecurity defenders for now.

The poster says Google shipped another frontier model soon after Argon that tops Fable, Opus and GPT-6 Astra on coding and cybersecurity, while cautioning the scores may be benchmark-maxxed until hands-on testing.

A developer congratulates Logan Kilpatrick, says Gemini 4 performs very well in their nanoGPT experiment, and questions why it isn't named 4.1 Argon if priced the same as Argon.

Polymarket reports Google unveiled Gemini 4 Carbon, its most powerful model yet, surpassing Claude Opus 5.5 and GPT-6 Astra on the Artificial Analysis Intelligence Index.

The post claims Gemini 4 Carbon outperforms Opus 5.5 on almost all benchmarks and is not yet public, with the author eager to try it.

A theory says Anthropic is delaying Fable 5.5 so it can release after GPT-6.1 Astra or Gemini 4 Carbon takes on Opus 5.5, then reclaim the top spot.

A user notes Gemini 4 Carbon launched just two weeks after Argon, scoring 59 on the new AA index to beat Fable and Astra, and looks forward to trying it.

Carbon scores 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71% on CWE-bench without guardrails, priced like Argon at $2/$10 and opening to selected cyber defenders this week.