Google has presented Gemini 4 Carbon, a multimodal model intended for cybersecurity tasks, which according to DeepLearning.AI took first place on the Intelligence Index list by Artificial Analysis, ahead of Anthropic's Claude Opus 5.5. As stated, the model is not yet available to the general public, but Google is providing it to a selected circle of cybersecurity companies, and a date for wider availability has not been announced.
According to the source, Carbon accepts text, images, speech and video up to one million tokens, and outputs text at a speed of 94.6 tokens per second. The model has adjustable reasoning and first place on Intelligence Index version 5.0 with a score of 59, second place on the Vals Index list with 70.36 percent and first place on the Arena.ai text ranking with 1,528 Elo points. The price via API is two, 0.10 and ten dollars per million input, cached and output tokens in the introductory period, and then four, 0.20 and 20 dollars. Parameters, architecture and training data have not been published.
Google states that Carbon continues the development of the Gemini 4 Argon model and that coding, science, multi-step agent tasks and multimodal understanding have been improved. The model is first on four of the nine individual Intelligence Index tests, including AutomationBench-AA for business tasks and AA-Omniscience for general knowledge. On the Coding Agent Index it takes second place behind Claude Opus 5.5, but completes tasks faster and at a lower price. On three demanding cybersecurity tests, CWE-bench, ExploitGym and SEC-bench Pro, it achieved 71.0, 54.8 and 88.2 percent, with these results being the company's own data.
According to the source, when the introductory discount expires, the price per token will be the same as with Claude Opus 5.5, but Carbon completes tasks with 29 percent fewer output tokens and 23 percent lower cost per task.
What this means for businesses
For companies considering the introduction of advanced models, the news indicates that competition is increasingly shifting from the price per token itself to efficiency per task. A model that uses fewer output tokens can significantly reduce operational costs in extensive agent processes, although the introductory price applies only for a limited time. Limited availability and the lack of data on architecture make independent risk assessment difficult, so for companies with sensitive data it is reasonable to plan pilot projects and comparative testing before wider application. The cybersecurity results are the company's own data and require additional verification.
This summary was written with AI based on the cited source and checked by automated rules. For exact details, read the original article.