Google's Gemini 4 Carbon Arrives With Strong Cyber Scores and a Staged Rollout
Google has announced Gemini 4 Carbon on October 28, 2026, calling it the strongest model in its line, with state-of-the-art cybersecurity and long-horizon capabilities (sustained, multi-step work on tasks that run for many steps rather than a single prompt). Google's announcement positions Carbon as the successor to Gemini 4 Argon in the Gemini 4 generation.
On version 5.0 of the Artificial Analysis Intelligence Index, Carbon ranks first with a score of 59. On WebDev Arena, which ranks models by head-to-head user votes converted into Elo ratings (the chess-derived scoring system where wins raise a rating and losses lower it), Carbon ranks fourth with 1769, after gaining 89 Elo points.
The cybersecurity figures are the centre of the announcement. Google reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching. Google also reports more than a 2x uplift in success rate for generating secure code over Argon on internal benchmarks, meaning tests Google ran itself rather than public ones.
Argon came first. Google said it launched Argon two weeks before the Carbon announcement. Reuters reported on September 30, 2026 that Alphabet had announced a new top-tier model to anchor the Gemini 4 generation, and on October 13, 2026 that Google had launched Argon for developers and businesses. Google describes Argon as its frontier model for real-world coding, enterprise knowledge work, and cyber defense. Its Gemini models page headlines Argon with the line "Built to solve your hardest problems." and states that Google is giving developers, enterprises, scientists, and consumers access to it.
Access to Carbon is staged. Google plans to make Carbon available starting this week to a subset of cyber defenders, then gradually extend access to additional trusted partners through its Fairwind Program. Google also said it is engaged in the U.S. government's voluntary process for pre-release model testing for Carbon.
For API customers, Google lists introductory pricing at $2 per million input tokens and $10 per million output tokens. Cached input tokens, which are prompt tokens the system has already processed and stored for reuse, are priced at 95% off the input token price.
In my view, the staged rollout matters as much as the benchmark table. A model that scores well at both finding and exploiting vulnerabilities is dual-use by construction, so deciding who receives it first has become a product and policy decision as much as a technical one. The defensive numbers make the stronger case for the release. A 71.0% patching score and a doubling of secure-code success rates, if they hold outside Google's test environment, could shorten the time that known weaknesses stay in production code. Defenders are the group most likely to turn that capability into fewer exposed systems, which is why their early access is the detail to watch.
The next signals will come from outside Google's own measurements: independent evaluations of the cyber benchmarks, how the Fairwind Program widens access, and whether the government pre-release testing process shapes the timing of broader availability.
