Google has announced Gemini 4 Carbon, a new frontier-class AI model that currently holds the top position on the Artificial Analysis Intelligence Index. The model, which features a 1 million token context window and multimodal capabilities for text, image, video, and speech, reportedly outperforms several competitors in generalized intelligence and knowledge recall. According to Artificial Analysis, the model leads the Intelligence Index V5 benchmark by one point over Anthropic's Claude Opus 5.5 and two points over OpenAI's GPT-6 Astra.
In addition to general intelligence, Gemini 4 Carbon is designed with a specific emphasis on cybersecurity and reliability. The model achieved an 88.2 percent score on the SEC-bench Pro vulnerability discovery benchmark and a 71 percent score on the CWE-bench security patches benchmark. Google reported that the model performs more than 10X better than another lab's frontier model when generating high-quality security patches, and it shows a 2x uplift in success rates for generating secure code compared to the company's previous Gemini 4 Argon model.
The model also claims a significant reduction in hallucinations, topping the AA-Omniscience v 1.1 test. Google stated that Carbon cuts unsupported claims in responses twice as much as the Argon model and reduces model overreach by 26% by presenting sources and suggested actions rather than pursuing its own intentions during uncertainties. Despite these advancements, the AA-Omniscience Index score of 54 suggests that even top-tier models in 2026 still face challenges regarding consistent information recall.
Google is currently limiting access to Gemini 4 Carbon through its Fairwind program, which provides the model to a subset of cyber defenders while the technology undergoes evaluation by the U.S. government. The company is following a deliberate rollout process to iterate on safety guardrails before making the model widely available to developers.
Pricing for the model is set at $4 per 1 million input tokens and $20 per 1 million output tokens, matching the standard pricing of Gemini 4 Argon. On the Artificial Analysis Intelligence Index, the model's cost per task is $7.76, which is approximately 23 percent less than Claude Opus 5.5 but 20 percent more than GPT-6 Astra. While the model leads in several categories, it falls behind Claude Code with Claude Opus 5.5 in programming and trails Grok 4.7 and Kimi K3.1 Pro in knowledge work.
