Table of Contents
Google has introduced Gemini 4 Carbon, a new artificial intelligence model designed to strengthen cybersecurity operations, improve software development, and handle complex tasks that require extended reasoning.
Announced on October 28, 2026, the model follows the release of Gemini 4 Argon earlier in the month. Google says Carbon delivers improvements in vulnerability detection, software patching, coding accuracy, and long-running AI workflows.
Despite its advanced capabilities, the company is taking a cautious approach to availability, initially limiting access to selected cybersecurity professionals.
Gemini 4 Carbon Takes the Lead in AI Benchmarks
According to Google, Gemini 4 Carbon achieved a score of 59 on version 5.0 of the Artificial Analysis Intelligence Index, placing it first at the time of the announcement.
The index evaluates advanced AI models across several measures of intelligence and performance.
Carbon’s ranking reflects Google’s efforts to improve its competitive position in the AI industry, where companies including OpenAI and Anthropic continue introducing increasingly capable models.
The release also comes just two weeks after Gemini 4 Argon became more widely available, demonstrating a faster development cycle for Google’s latest model family.
However, benchmark rankings can change as new models are released and evaluation methods evolve.
New Cybersecurity Capabilities Target Software Vulnerabilities
Cybersecurity is one of the main areas Google has prioritised in the new model.
The company reported strong results across three evaluations covering vulnerability discovery, exploitation testing, and software repairs.
| Cybersecurity benchmark | Reported score | What it measures |
| SEC-bench Pro | 88.2% | Finding software vulnerabilities |
| ExploitGym | 54.8% | Testing whether vulnerabilities can be exploited |
| Collinear’s CWE-bench | 71.0% | Developing fixes for security weaknesses |
Google stated that these evaluations were conducted without guardrails.
The results suggest that Carbon could assist security teams with identifying weaknesses, assessing their potential impact, and preparing corrective code.
Google also reported that its Chrome engineering team observed more than ten times better performance in producing high-quality patches compared with an unnamed competing frontier model.
Its Wiz security division reported promising results in web penetration-testing benchmarks.
These findings are based on the evaluations described by Google and do not guarantee equivalent performance in every real-world security environment.
Designed for Complex and Long-Running Engineering Tasks
Beyond cybersecurity, Gemini 4 Carbon is designed to work through complicated engineering assignments that involve multiple steps.
Google says the model can continue working on challenging problems over extended periods, including tasks involving large codebases and difficult software defects.
Such capabilities could help developers investigate problems, examine possible solutions, and prepare code changes across interconnected systems.
The model also recorded an 89-point improvement in its WebDev Arena Elo rating, reaching fourth place in the ranking reported by Google.
These developments indicate that Carbon is intended for broader software engineering work rather than cybersecurity alone.
Google Reports Improvements in AI Safety and Reliability
Google has also highlighted changes intended to improve the model’s behaviour during demanding tasks.
According to the company’s internal testing, Carbon produced approximately half as many unsupported claims as Gemini 4 Argon.
Google also reported a 0.0% attack success rate on Gray Swan’s indirect prompt-injection benchmark, meaning none of the evaluated attacks succeeded.
However, this result applies to that particular evaluation and does not mean the model is immune to all prompt-injection attempts.
The company additionally reported reductions in inappropriate actions and failures to complete authorised tasks.
These improvements are particularly relevant to AI systems that can perform extended workflows with limited human intervention.
Gemini 4 Carbon Will Reach Cyber Defenders First
Google is introducing the model through a phased release rather than making it immediately available to everyone.
The first group includes selected cybersecurity defenders who will receive access during the week of the announcement.
Availability will subsequently expand to additional trusted partners through Google’s Fairwind Program.
The company is also participating in the US government’s voluntary pre-release AI testing process.
Google plans to make Carbon available to developers, businesses, and consumers after further evaluation, but it has not announced a specific date for general access.
The restricted rollout reflects the need to evaluate powerful cybersecurity capabilities before wider deployment.
Gemini 4 Carbon Pricing and API Costs
Google has announced introductory token-based pricing matching Gemini 4 Argon.
| Usage | Introductory price | Price after introductory period |
| Input tokens | $2 per million | $4 per million |
| Output tokens | $10 per million | $20 per million |
Google also announced a 95% discount on cached input tokens relative to the applicable input-token price.
These rates describe the announced pricing structure and do not mean the model is already accessible to every developer.
The company has not specified the duration of the introductory pricing period in its announcement.
What the Launch Means for the AI Industry
The arrival of Gemini 4 Carbon adds another competitor to the market for advanced AI systems capable of handling software development and security operations.
Its focus on finding vulnerabilities and preparing fixes could be particularly relevant to organisations managing complex applications and infrastructure.
At the same time, the phased rollout means independent users will have limited opportunities to assess the model until access expands.
Real-world adoption will depend on reliability, cost, security safeguards, and performance outside controlled evaluations.
Conclusion
Gemini 4 Carbon represents Google’s latest effort to advance AI capabilities in cybersecurity and complex software engineering.
With strong reported benchmark results, improvements in long-running tasks, and additional safety measures, the model could become an important tool for developers and security professionals.
However, its practical impact remains to be established as Google gradually expands access and gathers feedback from early users.