Google Releases Gemini 4 Carbon Model With Cybersecurity Focus

Google has introduced Gemini 4 Carbon, its latest frontier artificial intelligence model designed for complex software engineering and cybersecurity defense.

The model arrives two weeks after the release of Gemini 4 Argon, reflecting rapid release cycles among leading machine learning research laboratories.

Key model details include a context window of 1M tokens, according to Artificial Analysis.

Google says it is participating in a voluntary U.S. government assessment process before wider distribution of Carbon.

Google executive Sundar Pichai presented the model in a public statement, while DeepMind leader Koray Kavukcuoglu described Carbon as the organization's strongest model to date.

Initial deployments will begin with a select group of cybersecurity defenders before access expands through Google's Fairwind program.

Google conducts background checks on organizations that apply to the Fairwind Program to verify security history and analyze their record of ethical operations, according to Google DeepMind.

Google DeepMind says its Fairwind programme has more than 650 partner organisations; this count refers to the wider programme, not to the limited Carbon launch cohort.

Benchmark Leadership and Evaluation Scores

Gemini 4 Carbon secured the top position on version 5.0 of the independent Artificial Analysis Intelligence Index with an overall score of 59.

In technical comparison evaluations reported by Google, Carbon captured the highest score across 15 of 17 tracked benchmarks against Argon, GPT-6 Astra, and Claude Opus 5.5.

MarkTechPost identified exceptions to Carbon’s lead: Claude Opus 5.5 ranked ahead on OSWorld-2.1, while GPT-6 Astra led the FrontierSWE test.

Google reported that Carbon gained 89 Elo points on WebDev Arena to reach fourth place, with a rating of approximately 1,770.

Google indicated that the system was built to handle long-horizon engineering tasks spanning multiple days, including tracing subtle software bugs across large codebases.

In workflow resilience testing, the architecture reduced task overreach by 26% and task under-service by 29% during complex executions.

Cybersecurity Defense Capabilities and Tooling

A central focus of the release centers on automated vulnerability discovery, penetration testing, and code remediation.

In Google-reported evaluations, Carbon scored 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench. The tasks concern bug discovery, exploitation and patching, respectively.

Google reported a more-than-tenfold improvement against an unnamed comparison model in an internal Chrome patching evaluation; the claim is not an independently replicated result.

Security platform Wiz also observed Carbon leading external benchmarks for web application penetration testing under its Scan for Good initiative.

Safety Alignment and Commercial Pricing Tiers

Internal safety evaluations showed that Carbon reduced unsupported factual claims by approximately half compared to the previous Argon release.

High-severity misalignment alerts in internal live coding environments declined by nearly 80%.

The model also recorded a zero percent attack success rate on the Gray Swan indirect prompt-injection benchmark while reducing cheating attempts on Terminal-Bench 4.0 to 0.4%.

Google announced introductory pricing of $2 per million input tokens and $10 per million output tokens for Carbon when broader API access becomes available.

Cached input tokens receive a 95% cost reduction, before standard developer rates later adjust to 4 dollars per million input tokens and 20 dollars per million output tokens.