Google Releases Gemini 4 Carbon, a Cybersecurity-Focused Model Just Two Weeks After Argon

Google has announced Gemini 4 Carbon, a new top-tier AI model that arrives only two weeks after Gemini 4 Argon. The company says Carbon makes big gains in coding, multi-step tasks and especially cybersecurity, and that it ranks first on the Artificial Analysis Intelligence Index (version 5.0) with a score of 59, ahead of Claude Opus 5.5 and GPT-6 Astra. For now, though, ordinary users cannot try it.

Who can use it, and when

Carbon is first going to a select group of cybersecurity defense companies, so they can test it before a broader release. Google plans to open it to more trusted partners through its Fairwind Program and later make it available widely. Google describes the model as “coming soon to developers, businesses, and consumers.”

That is the same rollout pattern Google used for Argon, which is now available in the Gemini app and in AI search mode. If the pattern holds, Carbon may eventually reach those places too, but Google has not given a date.

What Google says it can do

Google’s headline claims are about security work. The company says Carbon holds first place on several benchmarks, which are standardized tests used to compare AI models:

• SEC-bench Pro, which measures finding real-world bugs: 88.2%.
• ExploitGym, which measures turning a vulnerability into an exploit: 54.8%.
• CWE-bench, which measures patching problems: 71.0%.

Google also says the model is twice as likely as Argon to produce secure code, based on its internal benchmarks. In web development, it reports a jump of 89 Elo points to fourth place on WebDev Arena with a score of 1769. Elo is a rating system borrowed from chess that ranks models by head-to-head results.

Beyond raw scores, Google says Carbon is built for longer jobs. Its own teams have seen the model “run for days” through dead ends on complex engineering tasks, according to the company. These figures come from Google’s announcement and have not been independently confirmed in the material we reviewed.

Safety and honesty claims

Google also points to behavior changes it says make the model more dependable:

• Unsupported claims are cut by roughly half compared with Argon, based on Google’s internal testing.
• The model is said to resist prompt injection better. That is an attack where hidden instructions in a document or web page try to hijack an AI tool.
• When a request is unclear, Google says Carbon gathers sources and points out conflicts to the user instead of guessing.
• Overreach is down 26% and under-service is down 29%, meaning it should stay within the scope of a task more reliably.

Why it matters for readers

Most people will not notice Carbon directly yet. The practical point is the pace: Google is shipping major model updates in weeks rather than months, and the leading labs keep trading the top spot on benchmark tables. Benchmarks are useful signals, but they are chosen and reported by the companies themselves, so a first-place score does not guarantee a better experience in everyday apps.

The limited release also shows how seriously Google treats the security abilities of this model. A tool that is good at finding and exploiting software flaws can help defenders, so Google is starting with that group.

Pricing and what is still unknown

For developers, Google says introductory pricing will be $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off the input price. Afterward, the price rises to $4 per million input tokens and $20 per million output tokens. Tokens are the small chunks of text that AI services bill by.

Still unknown is the date for general availability, whether Carbon will appear in the Gemini app, and how it performs outside Google’s own tests. Independent reviews should help answer that once more people can use it.

Sources: Engadget, The Verge, 9to5Google