Google announces Gemini 4 Carbon AI model, but only select cybersecurity defense companies can use it for now

What happened

Google has announced Gemini 4 Carbon, its latest frontier artificial intelligence model, The Verge and Ars Technica reported. The Verge said the announcement came on Wednesday. Ars Technica noted it arrived just two weeks after Google rolled out Gemini 4 Argon.

Carbon is an improvement on Argon, according to Ars Technica, which also reported that it is not the Flash-style model the outlet had been waiting for. Like Argon, Carbon is getting a staged release that Ars Technica said could take a while.

Who can use Gemini 4 Carbon

For now, the model is available only to a select group of cybersecurity defense companies, The Verge reported. Google plans to eventually expand access to other partners in what it calls its Fairwind Program, according to The Verge.

Ars Technica pointed out that Google says Carbon is so capable that the company cannot let the public use it just yet. The outlet also noted that Google had taken a break from releasing new frontier models earlier this year before this recent run of launches.

Pricing

Carbon will launch at the same introductory price as Argon, The Verge reported. That is $2 per million input tokens and $10 per million output tokens.

What Google says the model can do

Google says Carbon reaches a new threshold in coding, agentic tasks and scientific analysis, Ars Technica reported. The Verge described Google's claims of "significant gains" in coding, agentic tasks and cybersecurity. Most of the performance figures below come from Google itself.

Cybersecurity claims

According to Ars Technica, Google says the model shows a "significant leap" in detecting, validating and fixing software vulnerabilities. Google also says Carbon is twice as likely as Argon to produce code free of vulnerabilities. The Chrome team claims Carbon is 10 times better at generating "high-quality" patches than non-Google models, Ars Technica reported.

Longer tasks and reliability

Google says Carbon is built to complete longer tasks, according to The Verge. The company said its internal teams have seen the model "run for days to persist through dead ends to tackle complex, multi-step engineering tasks." Google also says it is better at avoiding shortcuts, resisting prompt injection attacks and working with users.

Google says Carbon cuts "unsupported claims" by two times compared with Argon, The Verge reported. The company presents this as part of offering more "transparency and honesty" in the model.

How it ranks on benchmarks

Google highlighted Carbon's top spot on the Artificial Analysis Intelligence Index (v5.0), with Claude Opus 5.5 and GPT-6 Astra trailing behind, according to The Verge. The Verge also reported that Carbon holds the number one position on SEC-bench Pro, which measures how well an AI model discovers real-world bugs.

Carbon also holds the top spot on ExploitGym, a benchmark that measures how a model can turn a vulnerability into exploits, The Verge reported. Ars Technica said Google reported an 89-point increase on WebDev Arena, where Carbon now ranks number four on the leaderboard.

Both reports emphasize the model's strength at finding flaws in software and writing secure code. More coverage of artificial intelligence is available in our AI and tech section.

What to watch next

• How long Google's staged release of Carbon takes, which Ars Technica said could take a while.
• When access expands beyond cybersecurity defense companies to other partners in Google's Fairwind Program, as The Verge reported is planned.
• The launch pricing, which The Verge reported will match Argon's introductory price of $2 per million input tokens and $10 per million output tokens.