Google unveils Gemini 4 Carbon, claims top cybersecurity scores

Google announced Gemini 4 Carbon on Wednesday, October 28, calling it its strongest model, with “state-of-the-art cybersecurity” and long-horizon capabilities. Access starts this week with a limited group of cyber defenders.¹

Every performance number below comes from Google, not from independent testing.

Google says Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index, with a score of 59. The index maker unveiled that updated version the same day, with new coding and science benchmarks, and says 60% of it rests on private test sets or held-out answers.²¹

On cybersecurity, Google reports 88.2% on SEC-bench Pro, a test of finding vulnerabilities. It reports 54.8% on ExploitGym, which covers exploitation, and 71.0% on CWE-bench, which covers patching. Google calls this a firm lead. It also says Carbon showed lower misalignment and fabrication rates in live coding work with its own teams.¹

On web development, Google says Carbon gained 89 Elo points on WebDev Arena, landing fourth at 1768.¹ Ars Technica adds that Google says Carbon is twice as likely as its predecessor to produce code free of vulnerabilities. Google’s Chrome team claims it is 10 times better at producing high-quality patches than non-Google models.³

Carbon builds on Gemini 4 Argon, which launched two weeks earlier. Ars notes that Carbon beats Argon by slim margins on every test Google supplied, and that the rival models it was compared with, including GPT-6 Astra and two Claude models, sit only a few points behind. Ars also says it cannot tell whether Carbon is a wholly new model or a post-training upgrade of Argon.³

On safety, Google reports that Carbon resisted every attempted prompt-injection attack in the Grey Swan evaluation, a 0 percent success rate for attackers. The same test put Argon at 0.7 percent, Claude Opus 5.5 at 1 percent and GPT-6 Astra at 8.5 percent.³ A prompt injection is an attack in which malicious instructions are slipped into content a model reads, hoping it will obey them.

The rollout is staged. Google says it is taking part in the US government’s voluntary process for pre-release model testing. It will widen access to more trusted partners through its Fairwind Program, gathering feedback and adjusting guardrails before opening Carbon to developers, businesses and consumers “as soon as possible.”¹ Google has given no wide-release date.³

Pricing matches Argon’s introductory rate: $2 per million input tokens and $10 per million output tokens, with cached input tokens 95% cheaper.¹ VentureBeat reports the price later rises to $4 and $20.² Ars says Google would not name an end date for the introductory rate.³

Google also says Carbon can run for days on hard engineering tasks, such as tracking down and patching subtle bugs in large codebases and low-level systems.¹

Our read: The headline numbers are Google’s own, and Ars’ point stands that the leading models cluster within a few points of one another. The more telling detail is the rollout: a model Google says is this strong at finding and exploiting flaws is going first to defenders, not the public. Watch for independent results on the cyber benchmarks and for a date for wider access.

Google has promised to keep gathering feedback from early users and to improve its guardrails before a broader release.¹

Footnotes

1. Google Blog, “Gemini 4 Carbon: our next frontier model”. Retrieved October 29, 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

2. techmeme.com, “Google launches Gemini 4 Carbon model”. Retrieved October 29, 2026. ↩ ↩²

3. Ars Technica, “Google announces Gemini 4 Carbon, its latest frontier AI model that you can’t use”. Retrieved October 29, 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵