Google launched its Gemini 4 Carbon frontier artificial intelligence model on 28 October, designing the system specifically to automate complex software engineering and defend computer networks against vulnerabilities. The release builds directly on Gemini 4 Argon, which arrived two weeks earlier, and focuses on identifying code flaws and generating root-cause fixes.
Key points
- Gemini 4 Carbon leads general intelligence benchmarks and focuses heavily on vulnerability research and software patching.
- Google priced the model at an introductory $2 per million input tokens and $10 per million output tokens, undercutting rival frontier systems.
- Access is restricted initially to selected cybersecurity defenders and partner organisations under the Fairwind Program before a broader rollout.
- The Chrome security team reported the model produced fixes at ten times the success rate of a competing frontier system.
Benchmarking software defence
Google pitched the new system as its most capable AI release to date. While previous Gemini releases addressed general consumer and enterprise workflows, Gemini 4 Carbon targets technical operations that take software engineers hours or days to solve. The model specialises in navigating large codebases, isolating bugs in low-level systems and writing code patches that address flaws at the architectural level rather than applying surface-level workarounds.
In technical evaluations shared by Google, the system achieved an 88.2 percent score on SEC-bench Pro, an evaluation suite designed to test how well artificial intelligence spots flaws in production software. It reached 54.8 percent on ExploitGym, an environment testing vulnerability reproduction, and scored 71.0 percent on CWE-bench, which measures the ability of a system to fix known software weaknesses. Google said external testing partners conducted these evaluations without security guardrails.
Internal testing showed that the model generated secure code at more than twice the success rate of Gemini 4 Argon. The Google Chrome security team separately evaluated the model against another frontier AI system from an unnamed commercial lab, reporting that Carbon created high-quality software patches ten times more effectively. Google did not disclose the testing parameters or the identity of that competitor.
Agentic coding and general performance
Beyond defensive security, Google reported that Gemini 4 Carbon topped several industry benchmarks for general reasoning, agentic coding and autonomous computer operation. On the Artificial Analysis Intelligence Index version 5.0, Carbon ranked first with an overall score of 59, placing it ahead of Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Astra.
Writing in an announcement blog post, Google DeepMind executive Koray Kavukcuoglu, who serves as the company’s chief AI architect, noted that the model expands directly on the groundwork laid by Argon. The company reported that Carbon demonstrates improved resistance to prompt injection attacks, reduces shortcut-taking during multi-step logic workflows, and limits unsupported factual claims during extended automated tasks.
Web development benchmarks also shifted. On WebDev Arena, an industry leaderboard assessing full-stack code generation, Carbon moved into fourth place overall with a score of 1,769 points, representing an 89-point jump over previous Google models.
Pricing and competitive pressure
Google introduced the model with aggressive pricing intended to compete with Anthropic and OpenAI. The company set introductory API pricing at $2 per million input tokens and $10 per million output tokens. Those rates match the introductory pricing Google offered for Gemini 4 Argon earlier in October. Cached input tokens receive a 95 percent price reduction, lowering the cost of repeated queries against large codebases.
Standard pricing will later double, charging customers $4 for every million tokens fed into the model and $20 per million generated after the introductory window closes. Even after that adjustment, the pricing targets developer teams weighing the cost of running autonomous coding agents across multi-million-token repositories. Anthropic has faced heightened commercial scrutiny over inference expenses ahead of its planned initial public offering, creating an opening for Google to capture developer market share.
Phased rollout and security controls
Autonomous exploitation capabilities make cybersecurity models sensitive to release. Because the system can discover and evaluate software vulnerabilities, Google is limiting initial distribution rather than offering immediate open access.
The company is deploying the model first to pre-screened defensive cybersecurity firms and organisations participating in the United States government’s voluntary framework for pre-release frontier model evaluation. A wider pool of enterprise customers will gain access through Google’s Fairwind Program, an initiative designed to stress-test high-capability models in restricted environments.
Google followed an identical multi-stage deployment with Gemini 4 Argon, which eventually moved from restricted testing into consumer search products and developer tools. The company plans to bring Carbon to developers, corporate cloud customers and consumer Gemini applications once safety evaluations conclude, though it has not published a formal public launch date.
