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

The official GoogleAI account links to a blog post detailing Gemini 4 Carbon.

ValsAI reports Gemini 4 Carbon debuts at #2 on Vals Index v2.2 and takes first place on both its Finance Index and Legal Research Benchmark.

DeepMind lead Demis Hassabis posts a link to Google's official blog introducing the newly released Gemini 4 Carbon model.

Google says Carbon scores 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on Collinear's CWE-bench, with over 2x the secure-code success rate of Gemini 4 Argon on internal benchmarks.

Google launched Gemini 4 Carbon two weeks after Argon's debut, with benchmarks showing it beating GPT-6 Astra and Claude Opus 5.5 on most metrics.

With 60% private tests, Claude Opus 5.5 tops the new leaderboard at 58 while Gemini 4 Argon scores 50, just behind GPT-6.1 Terra at 51.

Google DeepMind confirms Gemini 4 Carbon as its next frontier model, state-of-the-art across cybersecurity and long-horizon capabilities.

Google's Logan Kilpatrick says Gemini 4 Carbon is a big jump over Argon from two weeks ago, SOTA on cyber and long-horizon tasks, with better usability in coding and improved alignment.

Sundar Pichai announces Gemini 4 Carbon, ranking #1 on Artificial Analysis Intelligence Index v5.0 at 59 and leading cybersecurity evals, with gradual access for trusted cyber defenders starting this week.

Running the same single-file HTML solar system prompt across four models on OpenRouter, Gemini 4 Argon finished fastest at 42 seconds and 3.3 cents and was the author's favorite.

Unconfirmed rumors say Gemini 4 Carbon may challenge GPT-6.1 Terra on coding and 3D tasks, while Flash targets speed and affordability.

In a Euro Truck Simulator recreation test, Gemini 4 Argon finished in 28 minutes versus 40 for GPT-6.1 Terra, and the author calls Argon the winner.