Google's Carbon AI model tops reasoning index for cyber defense

Topping the reasoning charts

Google has introduced Carbon, a new AI model designed for long-horizon engineering work. According to an independent evaluation by Artificial Analysis, Carbon achieved a score of 59 on the reasoning index at its high reasoning setting. This puts it at the top of the rankings, surpassing other major models like GPT-6 Astra and GPT-6.1 Sol.

The reasoning index is a comprehensive measurement that combines nine different evaluations. These tests cover a variety of complex skills, including scientific reasoning, coding, and general knowledge. The specific metrics used in this index include Terminal-Bench 4.0, FrontierCode 2.0, HLE-Diamond, and Terminal-Bench-Science 0.1.

Carbon's performance marks a significant step forward compared to Google's previous offerings. For instance, Google's Argon model scored 50 on the same version of the index. This nine-point gap highlights Carbon's specialized ability to handle more difficult, multi-step logic problems than its predecessor.

A tool for cyber defense

A major focus for Carbon is its performance in cybersecurity tasks. Google reports that the model scored 88.2% on SEC-bench Pro for discovering vulnerabilities and 71.0% on CWE-bench, which is a benchmark for defensive cybersecurity and AI evaluation. These tests were conducted with guardrails disabled, though Google notes that results may differ in the final safeguarded product.

In internal testing, the Chrome team found that Carbon was more than 10X better at producing high-quality patches than another unnamed large frontier model. Additionally, Google reports that Carbon more than doubles the success rate of Argon when it comes to generating secure code. These metrics suggest the model is highly effective at identifying and fixing software flaws.

The model also shows high resilience against specific types of attacks. On Gray Swan’s indirect prompt-injection benchmark, Carbon achieved a 0.0% attack success rate. This is a notable improvement over Argon, which previously led that specific benchmark with a 0.7% success rate. These results position Carbon as a robust tool for protecting digital infrastructure.

High costs and long tasks

While Carbon excels at complex reasoning, it requires significant computational resources to function. Artificial Analysis recorded 200 million output tokens for Carbon during its evaluation, which is about three times the 64 million median for comparable models. This high consumption results in a weighted average cost of $7.76 per task.

This price point places Carbon in the middle of the pack for high-reasoning models. It is more expensive than GPT-6.1 Sol, which costs $1.27 per task, but it is cheaper than Opus 5.5, which costs $10.04 per task. Google plans to offer starting prices of $2 for every million input tokens and $10 for every million output tokens before they eventually rise to $4 and $20, respectively.

The model is specifically built to handle 'long-horizon' work, meaning it can persist through dead ends on tasks that take days to complete. This is a departure from standard models that might struggle with multi-step engineering problems. However, users should note that the time to first token is currently high, with a median of 97.5 seconds on Google's API.

Restricted rollout and next steps

Google is not releasing Carbon to the general public immediately. Instead, it is following a restricted rollout for cyber-defenders. This process began with the Fairwind Program, which provides access to government authorities, critical infrastructure operators, and software maintainers.

Selected defenders will receive access to Carbon this week, with plans to expand further through the Fairwind Program to other trusted partners. This phased approach allows Google to gather feedback and refine the model's safety features. The company stated that broader access for businesses and consumers will follow once more work on guardrails is completed.

In addition to these partner programs, Google is participating in a voluntary process with the U.S. government for pre-release model testing. This oversight ensures that as the model moves toward a wider release, it meets specific safety and performance standards. For now, the model remains a specialized tool for high-stakes engineering and security environments.