Google announced Gemini 4 Carbon on October 28, expanding its Gemini 4 model family with improvements in coding, reasoning, scientific tasks and multimodal understanding.
Capabilities beyond cybersecurity
Carbon follows Argon. Google says it can sustain engineering work across large codebases for days, including diagnosing difficult bugs, and improves full-stack web development.
These agentic workflows involve completing a sequence of steps, using tools and responding to intermediate results. Multimodal understanding means processing more than one form of information, rather than text alone.
Google also reports fewer unsupported claims and better handling of ambiguous instructions in internal testing. These are company-reported findings.
What independent benchmarks show
Artificial Analysis lists Gemini 4 Carbon in its High configuration at 59 on version 5.0 of its Intelligence Index, the highest published score at the time of writing. The index combines nine evaluations across knowledge work, coding, science and general knowledge, with each category contributing equally.
The score measures performance across that particular collection of tests. Artificial Analysis notes that individual results can be more relevant when choosing a model for a specific task.
WebDev Arena lists Carbon fourth with a preliminary score of 1,768. Its displayed rank range spans second to sixth, reflecting uncertainty in the ranking. This evaluation focuses on web development, including workflows involving multiple steps and tool use.
Availability and initial access
Initial access is limited to selected cyber defenders, followed by additional Fairwind partners. Google plans broader availability for developers, businesses and consumers, without announcing a general-release date.
The Fairwind Program prioritizes organizations such as government agencies, healthcare providers and telecommunications operators. Applicants are vetted, and access cannot be resold or redistributed.
Token pricing
Introductory pricing is $2 per million input tokens and $10 per million output tokens. Google lists subsequent rates of $4 and $20, with cached input priced at a 95% discount.
Sources: Google — Gemini 4 Carbon announcement, Artificial Analysis — Intelligence Index v5.0, WebDev Arena — leaderboard
