Gemini 4 Carbon: Google Holds Back Its Strongest Model

Google's Gemini 4 Carbon leads on security benchmarks but goes to cyber defenders first. What it can do, API pricing and when you get it.

Gemini 4 Carbon: Google Holds Back Its Strongest Model
Image: Google

Google has a new top model, and you can't use it yet. Gemini 4 Carbon was announced on October 28, only two weeks after Gemini 4 Argon, and Google is holding it back on purpose while security teams test it first.

What happened

In Google's announcement, Koray Kavukcuoglu calls Carbon the company's strongest model, with the biggest gains in cybersecurity and long, multi-step tasks. Google says it scores 59 on version 5.0 of the Artificial Analysis Intelligence Index, which puts it first on that list. The security numbers are the headline:

  • Finding bugs: 88.2% on SEC-bench Pro
  • Proving they're exploitable: 54.8% on ExploitGym
  • Fixing them: 71.0% on CWE-bench

Google adds that Carbon more than doubles Argon's success rate at writing secure code in its own tests, and that internal teams have watched it keep working on hard engineering problems for days at a stretch. On safety, Google claims a 0.0% attack success rate on Gray Swan's indirect prompt injection benchmark and roughly half as many unsupported claims as Argon. These are all Google's own figures, so I'd treat them as a starting point until outside testers weigh in.

Who gets it, and what it costs

Access starts this week with a small group of cyber defenders, then widens to trusted partners through what Google calls its Fairwind Program. Google says it's also part of the US government's voluntary pre-release testing process. Developers, businesses and regular Gemini users will get it later, with no date given.

When the API opens, the introductory price matches Argon: $2 per million input tokens (roughly ₹190) and $10 per million output tokens (roughly ₹940). After the intro period it doubles to $4 and $20. Cached input is 95% off.

Why it matters

A model that can find, prove and patch software flaws cuts both ways. In defenders' hands it speeds up patching, and the staged rollout shows Google knows the same skills would help attackers too. We've already seen AI turn into a serious bug hunter this week, when Chrome 156 fixed 52 security bugs, many of them found by Claude. Carbon looks like Google's answer to that race.

The pace is worth noting as well. Two frontier releases two weeks apart is fast, even by this year's standards.

What to do

  • If you're a regular Gemini user, nothing changes for now. Carbon isn't in the app yet.
  • If you build on the Gemini API, plan your costs around the $4 and $20 rate, not the intro price, so you don't get caught out later.
  • If you run security for a company, keep an eye on the Fairwind Program. That's the only route in right now.