Google DeepMind says its next frontier model, Gemini 4 Carbon, is its strongest yet for cybersecurity and long-horizon work and it will roll out first to a small group of cyber defenders.
Update, Oct 28, 17:46 UTC · via Google Blog
Google Blog identified the author as Koray Kavukcuoglu, SVP, Google DeepMind and Chief AI Architect, Google.
Google Blog reported that "Our researchers and models are working hand in hand to advance capabilities faster, while strengthening our safeguards in tandem." The outlet also reported that "Carbon illustrates the pace at which we’re pushing the frontier forward."
That pitch landed Oct. 28, just two weeks after Gemini 4 Argon.
The sequence matters. Argon set the base for the Gemini 4 generation, then Carbon builds on it as a sustained-agent model, a system designed to run autonomously for days across multi-step tasks without losing the thread. Google introduced Argon two weeks earlier with a focus on hard engineering problems, and DeepMind now says Carbon extends that work specifically for finding, proving, and fixing flaws. Reddit chatter ahead of the announcement had already framed this as an iteration, with developers noting internal documents that identified a checkpoint called Barium-B as the public build of Argon and speculating Carbon was an upgraded checkpoint staying inside Gemini 4.x.
On external tests evaluated without guardrails, DeepMind reports Carbon at 59 on the Artificial Analysis Intelligence Index v5.0 for the top rank as of the announcement, plus 88.2% on SEC-bench Pro, a benchmark that measures autonomous discovery of real-world vulnerabilities, 54.8% on ExploitGym, a benchmark that tests whether a model can turn a disclosed flaw into a working exploit to validate risk, and 71.0% on Collinear's CWE-bench, which grades patching to root cause. In web development, Carbon added 89 elo points to reach 1769 for fourth place on WebDev Arena, which DeepMind ties to better output quality and artifacts.
Internal signals point the same way, though they are still company-run. DeepMind says Carbon more than doubled success rate for generating secure code over Argon on its own benchmarks, and cites the Chrome team reporting Carbon generated high-quality patches more than 10 times better than another lab's large frontier model. Wiz, its cloud security platform, reports Carbon leading on web penetration testing for its Scan for Good initiative. Those are defender-side gains, not attacker capability in the wild, and the benchmarks above were run without safety guardrails.
The rollout is deliberately narrow. DeepMind says it is engaged in the U.S. government's voluntary process for pre-release model testing and will make Carbon available this week to a subset of cyber defenders, then gradually expand to additional trusted partners through its Fairwind Program. Developers, businesses and consumers will get access via product surfaces as soon as possible after feedback and guardrail iteration. Pricing at launch matches Argon at $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off, before moving later to $4 input and $20 output when the introductory period ends.
Safety work is framed as paired with capability. Carbon builds on Argon's multi-layered stack of training, real-time misuse defenses, prompt-injection resilience and chain-of-thought monitoring, according to DeepMind. The company reports Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0% with attempts at 0.4% while improving accuracy, cut unsupported claims by about half versus Argon, trimmed overreach by 26% and under-service by 29% for more reliable multi-step execution, dropped high-severity misalignment alerts by almost 80% in live internal coding, and hit 0.0% attack success on Gray Swan's Indirect Prompt Injection Benchmark versus Argon's 0.7%.
What to watch is whether those lab and limited-field results hold once Carbon moves beyond the gated defender cohort and guardrails are on, and how quickly the Fairwind feedback loop turns into general availability.
