Google Launches Gemini 4 Carbon With 88.2% Vulnerability Detection and 71% Code Patching Scores at $2 Input and $10 Output per Million Tokens

Google has announced Gemini 4 Carbon, its latest frontier artificial intelligence model designed for advanced software engineering, complex reasoning, autonomous agentic workflows and cybersecurity defence. Developed by Google DeepMind, the model has achieved leading scores on several cybersecurity benchmarks and secured the top position on the Artificial Analysis Intelligence Index version 5.0 with a score of 59.

According to Google, Gemini 4 Carbon builds on Gemini 4 Argon, which was introduced two weeks earlier. The new model focuses on solving complex, multi-step engineering problems, identifying software vulnerabilities, validating security risks and developing code patches. Google said its internal testing also showed improvements in coding accuracy, model alignment and the reliability of task execution.

Gemini 4 Carbon Scores 59 on Artificial Analysis Intelligence Index

Gemini 4 Carbon ranked first on the updated Artificial Analysis Intelligence Index version 5.0, scoring 59. The model also recorded improvements in coding, agentic capabilities, scientific tasks, full-stack web development and multimodal understanding.

On WebDev Arena, Carbon gained 89 Elo points to reach fourth place with a rating of 1,768. Google attributed the improvement to better output quality and the ability to generate more effective web development deliverables.

The company said the model is designed to handle long-horizon engineering tasks that require sustained execution over extended periods. In internal testing, Carbon worked on complex problems involving root-cause analysis and bug fixes across large codebases and low-level systems, including tasks that required it to navigate multiple obstacles before reaching a solution.

Cybersecurity Benchmarks Show Improvements in Vulnerability Detection and Patching

Cybersecurity is a major focus of Gemini 4 Carbon, with Google reporting leading performance across vulnerability detection, exploitation validation and remediation.

The model achieved an 88.2% score on SEC-bench Pro, a benchmark focused on identifying vulnerabilities in real-world software. The result highlights its ability to locate complex security flaws that may be difficult to detect through conventional testing methods.

On ExploitGym, which evaluates the ability to validate vulnerabilities through exploitation, Carbon recorded a score of 54.8%. This capability can help cybersecurity professionals determine whether identified weaknesses present exploitable risks and prioritise remediation efforts.

For vulnerability remediation, Carbon achieved 71.0% on Collinear’s CWE-bench. The benchmark evaluates the model’s ability to address software weaknesses by generating code fixes that target their underlying causes.

Google stated that these results demonstrate improvements across multiple stages of the cybersecurity defence process, from discovering vulnerabilities to assessing their severity and implementing fixes. The benchmark evaluations cited for these results were conducted without guardrails.

Google Reports More Than 2x Improvement in Secure Code Generation

Google said Gemini 4 Carbon achieved more than twice the success rate of Gemini 4 Argon in generating secure code on its internal benchmarks.

The company’s Chrome team also reported that Carbon performed more than 10 times better than another AI laboratory’s large frontier model when generating high-quality patches in an internal comparison. Google did not identify the competing model in the announcement.

Wiz, Google’s cloud security platform, also observed strong performance from Carbon on web penetration-testing benchmarks. These capabilities are relevant to automated security assessments, vulnerability discovery and the development of more secure software systems.

The results are based on Google’s internal evaluations and reported comparisons. Performance in production environments may vary depending on the software, task complexity, security requirements and deployment conditions.

Gemini 4 Carbon Improves Reliability and Reduces Unsupported Claims

Google said it has made additional changes to Carbon’s training and safety systems to improve reliability during complex tasks.

In internal testing, the model reduced unsupported claims by approximately two times compared with Gemini 4 Argon. Google said Carbon is designed to communicate methodological limitations, identify missing evidence and explain which steps remain necessary rather than presenting unsupported conclusions.

The model also showed improvements in executing multi-step plans while respecting project requirements. According to Google, Carbon reduced overreach by 26% and under-service by 29%, indicating progress in balancing task completion with adherence to instructions and defined scope.

In live coding environments used by Google’s internal teams, high-severity misalignment alerts declined by nearly 80%.

Google also reported that successful cheating on Terminal-Bench 4.0 fell to 0%, while attempts to cheat declined to 0.4%. The company said these improvements were achieved alongside gains in overall accuracy.

These findings are based on internal testing and should not be interpreted as a guarantee that the model will never produce errors or behave unexpectedly in every real-world scenario.

Prompt Injection Resistance Reaches 0% Attack Success Rate

Gemini 4 Carbon also recorded improvements in resistance to indirect prompt injection, a security risk in which malicious instructions embedded in external content attempt to manipulate an AI system.

On the Gray Swan Indirect Prompt Injection benchmark, Carbon achieved a 0.0% attack success rate across the evaluated attempts. Google said Gemini 4 Argon had previously recorded a 0.7% attack success rate on the same benchmark.

The company said Carbon incorporates a multi-layered safety and security framework that combines model training, real-time misuse defences, prompt-injection resilience and chain-of-thought monitoring.

The improvements are intended to help the model operate more reliably in environments involving external data, software repositories and multi-step automated workflows. However, benchmark results reflect the specific attacks and evaluation conditions tested and do not establish immunity to all prompt-injection techniques.

Google Begins Limited Rollout Through Fairwind Programme

Google is adopting a phased deployment strategy for Gemini 4 Carbon, with initial access focused on a selected group of cybersecurity professionals.

Starting the week of October 28, 2026, the company plans to make the model available to a subset of cyber defenders before gradually expanding access to additional trusted partners through its Fairwind Programme.

Google also said it is participating in the US government’s voluntary pre-release model-testing process. The company intends to collect feedback from early users and refine its safeguards before making Carbon available more broadly.

Access for developers, businesses and consumers is planned for a later stage through Google’s product offerings. The company has not specified a definitive date for general availability in the announcement.

Gemini 4 Carbon API Pricing Starts at $2 per Million Input Tokens

Google said Gemini 4 Carbon will launch at the same introductory API pricing as Gemini 4 Argon.

The introductory pricing is as follows:

• Input tokens: $2 per million tokens.
•  Output tokens: $10 per million tokens.
• Cached input tokens: 95% below the standard input token price, equivalent to $0.10 per million cached input tokens if the full discount applies to the stated $2 rate.

The pricing structure provides lower-cost access to previously processed input content through caching, while output generation carries a higher per-token charge.

Actual API costs will depend on input and output token consumption, caching eligibility and the applicable pricing terms. The announced rates are introductory prices and may be subject to future changes.

Gemini 4 Carbon Targets Advanced Software Engineering and Cybersecurity

With Gemini 4 Carbon, Google is expanding its focus on AI systems capable of handling extended engineering workflows and cybersecurity tasks. Its reported benchmark results cover vulnerability discovery, exploit validation, code remediation and resistance to indirect prompt injection.

The model’s planned rollout will initially prioritise cybersecurity professionals and trusted partners, with wider access expected to follow after further testing and safety refinements.

The combination of advanced coding capabilities, long-horizon task execution and cybersecurity features positions Gemini 4 Carbon for applications involving complex software development and security operations. Its broader impact will depend on real-world performance, developer access and the effectiveness of its safeguards as deployment expands.