Gemini 4 Carbon leads by one point at Argon's price. Only cyber defenders can use it.

Google announced Gemini 4 Carbon on October 28 and called it its strongest model yet, built for cybersecurity and long-running agent work. It comes two weeks after Gemini 4 Argon opened to paid developers on October 13, and it launches at the same introductory price of $2 per million input tokens and $10 per million output tokens.

It also repeats Argon's rollout. This week Carbon goes only to a subset of cyber defenders. Google has not given a date for the Gemini API, the Gemini app or any Google AI plan.

What Google has confirmed

These details come from Google's announcement, written by Koray Kavukcuoglu, SVP of Google DeepMind and Google's Chief AI Architect, and from the benchmark table published with it.

• Access: a subset of cyber defenders gets Carbon this week. Google will then phase in more trusted partners through its Fairwind Program, and says developers, businesses and consumers will follow "as soon as possible" once it has gathered feedback and adjusted guardrails. Google is also taking part in the U.S. government's voluntary pre-release model testing.
• Introductory price: $2 per 1M input tokens and $10 per 1M output tokens, with cached input 95% off, the same as Argon.
• Standard price: $4 per 1M input tokens and $20 per 1M output tokens after the introductory period. Google has not said when that period ends. For comparison, Argon's $2/$10 rate runs through November 30, according to the Gemini API pricing page.
• Security benchmarks (run without guardrails): 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, each the top score among the models Google lists. On Gray Swan's indirect prompt-injection test, the attack success rate fell from Argon's 0.7% to 0.0%.

"Internally, our teams have seen Carbon run for days to persist through dead ends." - Koray Kavukcuoglu, Google DeepMind

Google's table compares Carbon with Argon, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 on 16 benchmarks, and Carbon is first on 14. The two it loses matter for developers. On FrontierSWE 2, a software engineering test, it scores 61.2% against Astra's 65.5% and Opus 5.5's 62.3%. On the OSWorld-2.1 computer-use test, it scores 81.5% against Opus 5.5's 87.2%. Its biggest margin is long context: 96.2% on GraphWalks at 256K to 1M tokens, against 71.8% for Astra and 66.8% for Opus 5.5.

What independent testers found

Artificial Analysis ranks Carbon first on its recently updated Intelligence Index v5.0 with 59 points, ahead of Claude Opus 5.5 at 58, GPT-6 Astra at 57 and Argon at 50. The lead is one point. These are v5.0 scores, so they do not line up with the older index numbers in our Argon launch coverage.

Other independent results, as reported by VentureBeat and The Decoder:

• Coding agents: on Artificial Analysis's Coding Agent Index, Carbon running in Google's Antigravity CLI scores 59. Claude Code with Opus 5.5 leads at 62.
• Accuracy and hallucination: on Artificial Analysis's knowledge test, Carbon answers 58% of questions correctly against Argon's 48%, but its hallucination rate rises from 4% to 11%.
• Cost per task: at the $4/$20 list price, Artificial Analysis puts Carbon at $7.76 per index task, close to Argon's $7.68. The Decoder reports $10.04 for Opus 5.5 and $6.25 for Astra, so Carbon comes in about 23% below Opus and about 24% above Astra.
• Professional tasks: on the Vals Index, Carbon ranks second at 70.36%, behind Claude Sonnet 5.5 at 70.85%. Sonnet 5.5 is not in Google's comparison table.

What is reported, not confirmed

Google has not said which product gets Carbon after the Fairwind phase. Varun Mohan, who leads Google Antigravity, wrote on X that "people have really enjoyed using it internally on Antigravity to solve their hardest tasks." According to VentureBeat, he indicated that access would widen through Fairwind before broader Antigravity availability, without a date. VentureBeat also notes that Artificial Analysis lists Carbon as having no publicly available API.

Several of Google's claims cannot be checked from outside yet: more than twice Argon's success rate at generating secure code on internal tests, a Chrome team report that Carbon wrote high-quality patches more than 10 times better than another lab's frontier model, and about half as many unsupported claims as Argon. These come from Google's own testing, and The Decoder notes they cannot be independently verified.

What this means for you

• If you use the Gemini app: nothing changes this week. Google AI Pro ($19.99 a month) and Ultra ($99.99 a month) include Argon, and Google has named no plan or date for Carbon. Our Gemini review, ChatGPT vs Gemini and Claude vs Gemini cover what you can subscribe to today.
• If you are testing Argon in the API: keep going. Carbon launches at the same introductory price and measured almost the same cost per task, so prompts and evaluations built on Argon should carry over when access opens. Budget at $4/$20 anyway, since Carbon's introductory period has no end date and Argon's ends November 30.
• If you code with Claude Opus 5.5 or GPT-6 Astra: there is no reason to move now. Claude Code with Opus 5.5 still leads the independent coding agent ranking, and Astra beats Carbon on FrontierSWE 2 in Google's own table. Carbon is the one to test first for long-context work once you can get it.
• If you run a security team: Fairwind is the only way in. Vulnerability finding and patching are Carbon's strongest results, and Google says it will add trusted partners gradually.

For timing, Argon went from defenders-only on September 30 to the paid API on October 13. Google has not promised the same gap for Carbon, so plan around the models you can buy today and treat Carbon as a test to schedule, not a switch to make. The model index has full specs for every model above.

Sources: Google announcement, Gemini API pricing, Artificial Analysis, Vals AI, VentureBeat, The Decoder, 9to5Google.