Gemini 4 Carbon: Google's Strongest Model Yet, and You Can't Use It Yet

Google announced Gemini 4 Carbon on 28 October, about two weeks after it opened Gemini 4 Argon to everyone. You can't use it yet. For now it is going to a small group of cyber defenders in Google's Fairwind Program, which is the same route Argon took before it went public.

What Google says it is for

Google calls Carbon its strongest model, with what it describes as state-of-the-art cybersecurity and long-horizon abilities. Long-horizon just means jobs with many steps that run for a long time. Google says its own teams have watched Carbon work on complicated engineering tasks for days. It also says the model makes about half as many unsupported claims as Argon, and oversteps what it was asked to do 26% less often.

The cyber numbers are the headline, and they are Google's own: 88.2% on SEC-bench Pro, which tests finding vulnerabilities, and 54.8% on ExploitGym, which tests exploiting them. Both are reported as new bests. That is also why access is narrow. A model that is this good at finding holes is good at using them, so Google is giving it to the people who defend systems first.

What independent testers found

Artificial Analysis, which runs its own tests, puts Carbon first on its Intelligence Index with a score of 59. Claude Opus 5.5 is on 58, GPT-6 Astra on 57 and Argon on 50. That is a lead of one point, so I wouldn't read much into the order at the top.

The picture is less one-sided on coding. On Artificial Analysis' coding-agent index, Carbon run inside Google's Antigravity CLI scores 59, but Claude Code with Opus 5.5 scores 62. Carbon is second on FrontierCode and on Terminal-Bench 4.0 (64%), and in Google's own table it trails on FrontierSWE 2 (61.2% against GPT-6 Astra's 65.5%) and on the offline OSWorld test (81.5% against Opus 5.5's 87.2%). Google says Carbon leads 15 of the 17 rows it chose to publish.

One result worth noticing: on Artificial Analysis' knowledge test, Carbon answers 58% of questions correctly, up from Argon's 48%, but its hallucination rate rises from 4% to 11%. It knows more, and it is also more willing to guess. That is the one place it does worse than the model before it.

The price, if you ever get it

Google doesn't say when Carbon's introductory price ends. Argon's ends on 30 November. Since nobody outside the Fairwind group can buy Carbon, the price is more of a signal than a number to plan around.

What to do in the meantime

If you are building on Argon, nothing changes. It is the public model, and Google has said nothing about retiring it. Google says Carbon will reach developers, businesses and consumers 'as soon as possible' and gives no date, and its post doesn't mention India. The sensible move is to keep the model name in a config setting rather than scattered through your code, so switching later takes a minute. Then wait for the first independent price and speed numbers before you decide whether it is worth paying for.

Sources

• Google: Gemini 4 Carbon, our next frontier model (28 October 2026)
• VentureBeat: Google unveils Gemini 4 Carbon (Artificial Analysis results)
• Google: Gemini 4 Argon is now available (13 October 2026)