Skip to content

AI

Gemini 4 Carbon lands two weeks after Argon. Pace, much?

Google's new frontier model takes the top spot on a fresh Artificial Analysis index and posts big security-benchmark numbers. It arrives just two weeks after its own predecessor, Gemini 4 Argon.

Cover photo: Grendelkhan / CC BY-SA 3.0

TL;DR

  1. Gemini 4 Carbon ranks #1 on Artificial Analysis Intelligence Index v5.0, scoring 59.
  2. Google cites 88.2% on SEC-bench Pro, 54.8% on ExploitGym, 71.0% on CWE-bench.
  3. Gemini 4 Argon shipped two weeks earlier. Release cadence is now a flex.

What happened

Google calls Carbon its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities. The headline claim: it sits at #1 on the freshly updated v5.0 of the Artificial Analysis Intelligence Index, with a score of 59.

The security numbers are the real pitch. Google says Carbon scores 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (exploitation) and 71.0% on CWE-bench (patching). It also claims a firm lead on those tests.

Google adds that Carbon shows significant gains in coding and agentic tasks. On internal benchmarks, it says secure-code generation improves, and that teams using it in live coding environments saw lower misalignment and fabrication rates.

Google's own framing: its researchers and models are working hand in hand to move faster. Read that sentence twice, then go refill your coffee.

Why it matters

A model that finds holes, simulates exploits and writes patches is a defender's dream. It is also, by definition, a tool that understands attacks. Google is selling the defense angle, and the benchmarks are its evidence.

Treat the numbers as vendor claims. These are Google's figures, published in Google's own announcement. Independent testing is what turns a slide into a fact.

The cadence is the other story. Argon, then Carbon, two weeks apart. If frontier releases now arrive at fortnightly speed, nobody's evaluation spreadsheet survives the quarter.

Two frontier-model releases in two weeks is the pace Google itself is pointing to.

What it means for Europe

Europe's angle is access. New frontier models from US labs have a habit of reaching EU users on a delay, or with features trimmed, as regulators and vendors sort out the AI Act. Whether Carbon follows that pattern is not yet clear.

For European security teams, the question is practical: can you actually use it, and under what terms? A model strong at vulnerability work will draw regulatory attention precisely because of what it can do.

What to do with it

  • Check the benchmarks yourself.Read the SEC-bench Pro, ExploitGym and CWE-bench results before repeating Google's numbers in a slide deck.
  • Pilot it on defense work.If you have access, try vulnerability triage and patch drafting on a non-critical codebase first.
  • Verify availability in your region.European teams should confirm access and terms before planning a rollout around Carbon.
  • Track the index.Watch the Artificial Analysis Intelligence Index v5.0, since rankings may shift again within weeks.

Keep up, sapiens. The models already did, and they did it twice this month.