- Google announced Gemini 4 Carbon, a model specialized in cybersecurity, coding and complex multi-step tasks.
- Google said Carbon ranked first in the Artificial Analysis Intelligence Index v5.0 with 59 points.
- The company reported scores of 88.2% for vulnerability discovery, 54.8% for exploitability testing and 71% for patch generation.
Google announced the release of Gemini 4 Carbon, a model specialized in cybersecurity, coding and executing complex multi-step tasks. The company said the model ranked first in the updated Artificial Analysis Intelligence Index v5.0 with a score of 59 points.
Carbon is built on Gemini 4 Argon, which Google introduced two weeks earlier. The developers said Carbon improves agentic workflows, code generation and work on large software projects.
According to Google, Carbon scored 88.2% on SEC-bench Pro, which measures vulnerability discovery in real-world software. It recorded 54.8% on ExploitGym, a test of whether vulnerabilities can be exploited, and 71% on CWE-bench, which evaluates the generation of fixes addressing vulnerabilities’ root causes.
Google also said Carbon more than doubled its safe-code generation success rate compared with Argon in internal tests. According to the Chrome team, Carbon performed more than 10 times better than an unnamed competing frontier model when producing high-quality fixes.
In web development, Carbon gained 89 ELO points and ranked fourth on WebDev Arena with a score of 1,768, Google said.
Safety and reliability testing
Google said Carbon became more reliable when handling complex tasks and operating under uncertainty. In the company’s internal assessments, unfounded claims fell by about half compared with Argon, while alerts about critical inconsistencies in internal teams’ workflows declined by nearly 80%.
On Terminal-Bench 4.0, Carbon recorded a 0% rate of successfully bypassing tasks to take shorter paths to a result, while relevant attempts accounted for 0.4%. The model also reduced excessive actions by 26% and under-execution of tasks by 29%, according to Google.
In the Gray Swan Indirect Prompt Injection test, Carbon blocked all evaluated prompt-injection attack attempts. Its attack success rate was 0%, compared with 0.7% for Argon.
Google said the external cybersecurity benchmark results were obtained without the model’s protective guardrails enabled. The company is continuing to refine Carbon’s safety mechanisms before a broader rollout.
Access and pricing
Google will initially make Carbon available to a limited group of cybersecurity specialists. It plans to expand access later through the Fairwind program for trusted partners.
The company also said it is participating in the U.S. government’s voluntary pre-deployment model testing process. After further testing and safety improvements, Google plans to make Carbon available to developers, businesses and everyday users.
Gemini 4 Carbon will start at the same rates as Argon: $2 per million input tokens and $10 per million output tokens. Google will offer a 95% discount on input tokens for cached text.
Google previously created a crypto-asset protection system for banks.
Source: Incrypted
