Google has introduced Gemini 4 Carbon, a new artificial intelligence model designed to handle complex coding projects, advanced reasoning, and cybersecurity tasks. The model has taken the top position on the Artificial Analysis Intelligence Index version 5.0, placing it ahead of several leading AI systems.
Announced on October 28, 2026, Gemini 4 Carbon received an intelligence score of 59 points. Artificial Analysis independently lists the model at number one on its index, with Claude Opus 5.5 and OpenAI’s GPT-6 Astra close behind. Source: Artificial Analysis
The release marks another development in competition among Google, OpenAI, and Anthropic. However, Google is initially limiting access to selected cybersecurity professionals and trusted partners rather than opening the model to everyone.
Gemini 4 Carbon Leads the Artificial Analysis Intelligence Index
Gemini 4 Carbon scored 59 on the updated Artificial Analysis Intelligence Index. Claude Opus 5.5 scored 58, while GPT-6 Astra received 57. Google’s previous model, Gemini 4 Argon, scored 50 on the same version of the index.
The results place Carbon at the top of this particular evaluation, which compares advanced AI models across several capabilities. However, the overall ranking does not mean that Carbon performs best on every individual test.
| AI model | Intelligence Index score |
|---|---|
| Google Gemini 4 Carbon | 59 |
| Anthropic Claude Opus 5.5 | 58 |
| OpenAI GPT-6 Astra | 57 |
| Google Gemini 4 Argon | 50 |
The small difference between the leading models highlights how closely major AI developers are competing. Performance can vary depending on the benchmark, the task, and the model’s reasoning settings.
Google Reports Improvements in Coding and Complex Tasks
Google says Gemini 4 Carbon is designed to work through demanding engineering problems that require multiple steps and sustained effort. These include identifying the underlying causes of software bugs, modifying large codebases, and developing fixes for complicated technical issues.
According to Google’s published benchmark comparisons, Carbon scored 65.2% on Terminal-Bench 4.0, ahead of Claude Opus 5.5 at 59.6% and GPT-6 Astra at 59.1%.
The model also achieved 77.4% on Agents’ Last Exam, a benchmark included in Google’s comparison of AI capabilities. Its performance reflects improvements in tasks that involve using tools and completing sequences of actions.
Nevertheless, Carbon does not lead every coding evaluation. GPT-6 Astra scored 65.5% on FrontierSWE 2, compared with Carbon’s 61.2%. Claude Opus 5.5 also performed better on the OSWorld-2.1 computer-use benchmark, scoring 87.2% against Carbon’s 81.5%.
These differences show why developers should compare models against their specific requirements instead of relying on a single overall score.
Gemini 4 Carbon Targets Cybersecurity Challenges
Cybersecurity is one of the main areas Google has highlighted for Gemini 4 Carbon. The company says the model can assist security professionals with finding software vulnerabilities, evaluating weaknesses, and developing patches.
Google reported that Carbon achieved the following results on cybersecurity benchmarks:
- SEC-bench Pro: 88.2% for identifying software vulnerabilities.
- ExploitGym: 54.8% on a benchmark evaluating vulnerability exploitation.
- CWE-bench: 71.0% for addressing software weaknesses.
Google also reported more than twice the success rate of Gemini 4 Argon for secure-code generation on its internal benchmarks. These are company-reported findings and should not be treated as a guarantee of performance in every real-world security environment.
The model’s cybersecurity capabilities could be useful to security teams working to identify weaknesses and improve software protection. Its deployment, however, remains controlled as Google continues evaluating risks and safeguards.
Google Reports Better Safety and Reliability
Alongside its technical improvements, Google says it has worked to make Carbon more reliable during complex tasks. The company reported fewer unsupported claims than with Gemini 4 Argon and improvements in how the model follows project requirements.
Google also said Carbon recorded a 0.0% attack success rate on the Gray Swan Indirect Prompt Injection Benchmark under the reported test conditions. This result indicates that the evaluated attacks were unsuccessful during that assessment, but it does not prove that the model is immune to all prompt-injection attacks.
The company is continuing to gather feedback and refine safeguards during the initial rollout. Wider availability will depend on its deployment approach and ongoing safety work.
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 compared with the input-token price.
After the introductory period, the listed standard prices will be $4 per million input tokens and $20 per million output tokens. Google has not provided a confirmed end date for the introductory rates.
Access is initially being offered to a selected group of cybersecurity defenders. Google plans to expand access gradually to additional trusted partners through its Fairwind Program before making Gemini 4 Carbon available to developers, businesses, and consumers through its products.
As of October 29, 2026, Google has not announced a specific date for general public availability.
What Gemini 4 Carbon Means for the AI Industry
Gemini 4 Carbon’s leading score on the Artificial Analysis Intelligence Index adds to the competition between Google, OpenAI, and Anthropic. Its reported gains in cybersecurity and complex engineering tasks may be particularly relevant to organizations that use AI for software development and security work.
Still, the benchmark results show a mixed picture across different tasks, with competing models retaining advantages in some areas. Real-world usefulness will depend on reliability, cost, safety, and performance in the applications people actually use.
In other news read more about: Google Gemini 4 AI Agents Make History by Beating Minecraft Without Human Assistance
For now, Gemini 4 Carbon has secured the top position on one major AI ranking, while its limited rollout means most users will need to wait for broader access.
