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

The post checks off Gemini 4 Carbon over Opus 5.5 and leaves Gemini 4 Flash (Hydrogen) versus GPT 6.1 Sol as the next item to verify.

A user says Gemini Carbon ranks 2nd on Agent Arena versus 7th for Argon, and that hands-on time with Carbon before release felt much better than Argon, calling it a real deal from Google.

The post says Google just released Gemini 4 Carbon, which leads across multiple benchmarks spanning long-context reasoning, computer use and cybersecurity.

The poster says Gemini 4 Carbon scores 59 on Artificial Analysis Intelligence Index v5.0 versus Opus at 58, and that an Argon checkpoint on Arena routing to Carbon impressed in testing.

The poster contrasts Gemini 4 Argon (Oct 13) with Carbon (Oct 28), saying Google moved from parity to leadership and praising its AA score.

At the same $4/$20 standard pricing, Gemini 4 Carbon costs $7.76 per Intelligence Index task, roughly in line with Argon's $7.68 while scoring 9 points higher.

A chart shows Gemini 4 Carbon gaining 9 points over Argon, which Google released just two weeks prior, highlighting the rapid iteration pace.

A post says Carbon outperforms Opus 5.5, Fable 5.1 and GPT-6 Astra, noting Argon shipped only 15 days earlier and Google now sets the pace.

Google's blog claims internal teams have seen Carbon run for days through dead ends to root-cause and patch subtle bugs across large codebases and low-level systems.

Arena reports Gemini 4 Carbon (High) on the Pareto frontier for Agent, Code WebDev and Text, ranking #2 in Agent Arena at $1.80 median cost per task, 72% below Claude Opus 5.5 (Max).

DeepMind released Gemini 4 Carbon two weeks after Argon, scoring 59 on Artificial Analysis Intelligence Index v5.0, working autonomously for days, finding 88.2% of vulnerabilities and fixing 71.0%, initially limited to select cyber defenders.

A user reacts to another Google model, Carbon 4, and wonders if Google is scrambling because Argon 4 underdelivered.