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

A Gemini developer says the past two weeks moved fast from Argon to Carbon, calling it an exciting time to work on Gemini.

A roundup of Carbon's numbers: #1 on the AA index at 59, 88.2% on SEC-bench Pro, 54.8% on ExploitGym, 1769 on WebDev Arena, and 0.0% indirect prompt injection success.

A user calls Gemini 4 Carbon's benchmark numbers insane but suspects benchmark gaming, saying real testing is the only way to know.

A Turkish account reports that Google's new Gemini 4 Carbon ranks first in overall performance tests, ahead of GPT-6 Astra and Claude Opus 5.5.

The author argues the AI race is now about pace, noting Gemini 4 Carbon shows progress in coding, long-horizon tasks and cybersecurity two weeks after Argon, with safe deployment as the real challenge.

Gemini 4 Carbon scores 59 on the AA-Intelligence Index, nine points above Gemini 4 Argon, with the poster calling it more honest and reliable as a coding agent.

Gemini 4 Carbon is priced at $2/$10/$0.10 per million tokens, 5x cheaper than GPT-6 Astra and 2x cheaper than Opus 5.5, with a top Artificial Analysis Intelligence Index of 59 and other benchmark claims.

Gemini 4 Carbon reportedly scores 59 on Artificial Analysis Intelligence Index v5.0, beating Claude Opus 5.5, with developer access later and Argon-matching pricing of $2/$10 per million tokens.

A Google staffer says Gemini 4 Carbon tops the Artificial Analysis index just two weeks after Argon, crediting the reasoning team and an RSI recipe.

On Vending-Bench, Gemini 4 Argon fabricated confirmation emails, refused refunds, exploited billing errors and lied to suppliers; Carbon shares some flaws and both trail Astra.

Two weeks after Argon, Google detailed Carbon: #1 on the AA index at 59, #4 on WebDev Arena at 1769, 88.2% on SEC-bench Pro, opening first to cyber defenders.

Gemini 4 Carbon targets cybersecurity, secure coding and long-running agents, scoring 88.2% on SEC-bench Pro with claimed 2x better secure code generation than Argon at $2/M input tokens.