Google’s New AI Model Takes Aim at Cybersecurity’s Hardest Problems

Google has unveiled Gemini 4 Carbon, a new frontier AI model built for cybersecurity, complex software engineering, and long-running coding work. The model leads 15 of 17 benchmarks on Google’s charts, yet most developers and businesses cannot use it today.

That tension defines Carbon’s launch. Google claims a major jump over Gemini 4 Argon, while keeping the same introductory token prices. But access begins with selected cybersecurity defenders, and Google has not given a date for wider availability.

A Benchmark Leader With Rivals Close Behind

Independent evaluator Artificial Analysis places Carbon first on its Intelligence Index v5.0 when using high reasoning, with a score of 59. Claude Opus 5.5 follows at 58, GPT-6 Astra reaches 57, and Argon sits at 50. Carbon’s score improves by 9 points over Argon from four weeks ago.

Google’s own comparison shows Carbon improving on Argon in all 17 reported benchmark rows, with the highest result in 15 of them. OpenAI retains an advantage on one software engineering test, while Anthropic leads on one computer-use evaluation.

Carbon’s biggest gains over Astra and Opus appear in tasks that matter to companies building AI agents and developer tools. It leads by 24 points on long-input reasoning, 12 points on hacking, 11 points on workflow automation, and 8 points on math.

GraphWalks shows the scale of the gap across contexts ranging from 256,000 to 1 million tokens. Carbon reaches 96.2%, compared with Astra’s 71.8% and Opus’s 66.8%.

Google also says Carbon marks a “significant leap” in vulnerability detection, validation, and remediation. Carbon is twice as likely to generate code free of vulnerabilities compared with Argon, while the Chrome team claims Carbon produces “high-quality” patches 10 times better than non-Google models.

On WebDev Arena, Carbon gained 89 points and now ranks number four on the leaderboard. In the Grey Swan evaluation, it resisted every attempted prompt-injection attack, recording a 0 percent score compared with Argon at 0.7 percent, Opus 5.5 at 1 percent, and GPT-6 Astra at 8.5 percent.

Built for Work That Runs for Days

Carbon targets software projects that demand more than a single answer. Google says its agents have assisted with memory optimization and C/C++ migrations to Rust, with sensitive code changes passing through testing and human review.

“Internally, our teams have seen Carbon run for days,” Google wrote in the announcement. Varun Mohan, the former Windsurf CEO and current Google Anitgravity head, also said that people have enjoyed using Carbon internally on Antigravity to solve their hardest tasks.

Cybersecurity sits at the center of the release. Argon was designed to discover flaws, validate their significance, and produce fixes, while Carbon focuses on extending that work through longer coding and reasoning tasks.

Independent testing also reveals tradeoffs. Carbon improves over Argon in coding and knowledge work, but trails several rivals on professional tasks and records a higher hallucination rate than Argon on a factual-knowledge test.

Carbon’s announcement does not provide an API specification equivalent to Argon’s documentation. Argon supports more than 1 million tokens of shared context across input and output, with extended generation beyond its default 65,536-token output limit. It accepts text, images, video, audio, and PDFs, generates text, and supports function calling, code execution, structured responses, and computer interaction in preview.

Limited Access and a Temporary Price Advantage

Google is granting initial access only to selected cybersecurity defenders through its curated Fairwind Program. The company says broader access will come “as soon as possible” after it gathers feedback and iterates on guardrails.

Google is working within the voluntary testing framework established by the US government and plans to roll Carbon out through its Fairwinds cyber defense program once that work is done. The announcement promises controlled expansion but gives no product-by-product schedule or public release date.

Argon offers a useful comparison. Google introduced it on September 30, 2026, gave approved cybersecurity partners access through Fairwind, and expanded availability on October 13. Carbon arrives 15 days after that expansion. If Carbon follows a similar path, it could reach additional users within a couple of weeks, but Google has made no such promise.

Varun Mohan said access would expand through Fairwind ahead of wider Antigravity availability, but his post supplied no release date. Carbon will likely reach AI Ultra subscribers and Gemini API developers before a wider rollout, though that sequence remains unconfirmed.

Pricing gives businesses another reason to watch the launch. Google will charge $2 per million input tokens and $10 per million output tokens during the introductory period, with a 95% discount for cached inputs. Those rates are half of Opus 5.5’s and one-fifth of Astra’s.

The discount will not last. A footnote lists subsequent prices of $4 per million input tokens and $20 per million output tokens, doubling Carbon’s introductory rates. Google has not specified when the introductory period ends.

For an illustrative workload using 10 million uncached input tokens and 1 million billed output tokens, the introductory rates create a $30 token bill. The later rates produce $60, excluding other charges and assuming unchanged token use.

Carbon’s output charges include reasoning tokens, and its low, medium, and high reasoning settings affect consumption and response time. Google also bills tokens tied to internal tool instructions and function definitions when grounding is used, while Google Search grounding carries separate query charges.

The Model Is Powerful, But Still Behind the Gate

Carbon’s 95% cached-input discount stands out, but Google has not explained its caching mechanics. Argon’s guide describes automatic caching and says explicitly created caches are unsupported, so developers cannot assume Carbon works the same way.

For now, Argon remains the public model customers can evaluate through its existing offering, including the Gemini app and AI search mode. Its API became available on October 13, alongside access through paid consumer products, AI Studio, and Android Studio.

Carbon brings stronger benchmark results, deeper cybersecurity ambitions, and long-running agent work to Google’s model lineup. The next test is not only whether it can top the charts, but whether Google can turn a tightly controlled preview into a dependable tool for developers and businesses.

Stay connected via Google News

Follow us for the latest AI News