Google has announced Gemini 4 Carbon, a new artificial intelligence model that has taken the top position on the Artificial Analysis Intelligence Index, outperforming several competing frontier models in the latest evaluation.
Carbon scored 59 on version 5.0 of the Artificial Analysis Intelligence Index, placing it ahead of Claude Opus 5.5, which scored 58, and OpenAI’s GPT-6 Astra, which scored 57 in the comparison cited by the ranking.
Google announced the model on October 28, 2026, describing it as its strongest AI model yet for complex coding, autonomous AI agents and cybersecurity.
However, access remains restricted during the initial rollout, with the company first making it available to a selected group of cybersecurity defenders before expanding access to other users.
The independent Artificial Analysis evaluation also identifies Carbon as the leading model on its Intelligence Index, although its performance varies across individual tasks and benchmarks. The ranking measures a defined set of capabilities rather than every possible use of an AI model.
Gemini 4 Carbon Leads the Latest AI Intelligence Ranking
According to Artificial Analysis, Carbon’s score of 59 puts it at the top of the updated Intelligence Index. The results reflect improvements in reasoning, scientific tasks, knowledge-based evaluations and coding-related work.
Google has also reported strong results across several specialist benchmarks. Carbon achieved 88.2% on SEC-bench Pro, which evaluates vulnerability detection, 54.8% on ExploitGym for exploitation-related tasks, and 71.0% on CWE-bench for vulnerability remediation.
These results point to potential applications in software development and defensive cybersecurity. However, Google’s benchmark claims should be considered alongside the testing methodology, evaluation conditions and results from other independent assessments.
Carbon does not lead every coding or computer-use test. The distinction matters because benchmark rankings can vary according to the task being measured, the model configuration and the tools available during evaluation.
Designed for Complex Coding and Autonomous AI Agents
Google says Gemini 4 Carbon is designed to handle long-running engineering tasks that require sustained reasoning and multiple steps.
According to the company’s official announcement, internal teams have used the model on challenging projects involving complex bugs, large software repositories and low-level systems.
These workflows require an AI model to investigate a problem, test possible solutions, recover from unsuccessful attempts and continue working toward a result. Such capabilities are increasingly important for AI agents that can perform tasks across software development environments with limited human intervention.
Google also reported an improvement of 89 Elo points on WebDev Arena, bringing Carbon to a reported score of 1,769 in the company’s announcement. The company said the gains reflect improvements in full-stack web development and the quality of generated outputs.
The reported results suggest that Google is targeting professional software engineering and other complex workflows rather than limiting Carbon to conventional chatbot interactions.
How Gemini 4 Carbon Compares With GPT-6 Astra and Claude Opus 5.5
Google’s published benchmark comparisons show Carbon ahead of several competing models on a number of coding, reasoning and computer-use evaluations.
On Terminal-Bench 4.0, Carbon scored 65.2%, compared with 59.6% for Claude Opus 5.5 and 59.1% for GPT-6 Astra. On Agents’ Last Exam v1.1, which evaluates computer-use capabilities, Carbon recorded 77.4%, ahead of GPT-6 Astra at 73.1% and Claude Opus 5.5 at 71.6%.
Carbon also performed strongly on GraphWalks, a benchmark for reasoning across long contexts. Google reported a score of 96.2%, compared with 71.8% for GPT-6 Astra and 66.8% for Claude Opus 5.5.
However, the results are not uniformly in Carbon’s favour. GPT-6 Astra recorded 65.5% on FrontierSWE 2, ahead of Carbon at 61.2%, while Claude Opus 5.5 led OSWorld-2.1 with 87.2%, compared with Carbon’s 81.5%.
These differences show why developers and businesses should assess models against their actual requirements rather than relying exclusively on an overall intelligence ranking.
Google Highlights Cybersecurity and AI Safety Improvements
Cybersecurity is one of the main areas Google has emphasised in the Gemini 4 Carbon announcement.
The company says the model can help security teams identify software vulnerabilities, assess whether flaws can be exploited and develop patches to address the underlying problems.
Google reported that Carbon more than doubled Argon’s success rate in internal secure-code-generation tests. The company also said its Chrome team observed substantially better patch-generation performance than another large frontier model in its internal evaluation.
These findings are company-reported results and should not be interpreted as a guarantee that AI-generated code will be secure in production environments. Human review, independent testing and established security procedures remain important when using AI in sensitive systems.
Google also reported improvements in resistance to prompt-injection attacks, reduced unsupported claims and fewer high-severity misalignment alerts during internal testing. On the Gray Swan Indirect Prompt Injection Benchmark, the company reported a 0% attack success rate in the evaluated test.
A zero score in a particular benchmark does not establish that a model is immune to all attacks. Real-world security depends on the application, deployment environment, safeguards and the nature of the threats encountered.
Gemini 4 Carbon Pricing and Availability
Google has announced introductory API pricing of $2 per million input tokens and $10 per million output tokens. Cached input tokens receive a 95% discount against the regular input price.
After the introductory period, the announced standard rates are $4 per million input tokens and $20 per million output tokens. The introductory offer is temporary, and users should check the latest pricing terms before deploying the model at scale.
At launch, Google is providing access to a selected group of cybersecurity defenders. It plans to expand access through its Fairwind Program before making Carbon available more broadly to developers, businesses and consumers.
The company has also said it is participating in the US government’s voluntary pre-release model testing process as part of its rollout strategy.
For now, Gemini 4 Carbon’s immediate significance lies in its reported performance on independent intelligence evaluations and its focus on long-running coding and cybersecurity work. Its wider practical impact will become clearer as access expands and developers gain more opportunities to evaluate its capabilities, reliability and cost on real-world tasks.
