Gemini 4 Carbon is the latest frontier model from Google DeepMind, announced on October 28, 2026. The official blog presents it as the company's strongest model to date, focused on cybersecurity and long-horizon agent work. It ranks first with a score of 59 on the updated Artificial Analysis Intelligence Index v5.0, succeeding Gemini 4 Argon, which launched two weeks earlier, as Google's flagship frontier model.
Leading on three security benchmarks, with test conditions that matter
All three published scores come from the cybersecurity defense pipeline: 88.2% on SEC-bench Pro for finding real vulnerabilities, 54.8% on ExploitGym for validating exploitability, and 71.0% on CWE-bench for patching at the root cause, each the best result on its benchmark at the time. DeepMind notes that these external benchmarks were evaluated without guardrails, so performance inside real products, where safety layers apply, should not be treated as identical.
In internal evaluations, Carbon more than doubled Argon's success rate at generating secure code. The Chrome team reports it performed more than 10 times better than another lab's frontier model at generating high-quality patches, and the cloud security platform Wiz also observed it leading on web penetration testing benchmarks. On coding, it gained 89 Elo points on WebDev Arena to reach fourth place, and Google says it can work for days on debugging and patching tasks spanning large codebases.
Cyber defenders first, priced the same as Argon
The rollout is deliberately slow: starting this week, access goes first to a subset of cyber defenders, then gradually expands to more trusted partners through the Fairwind Program. Google says it is taking part in the U.S. government's voluntary pre-release testing process, and will move as soon as possible to developers, businesses, and consumers after gathering feedback and iterating on guardrails. Pricing matches Argon at 2 dollars per million input tokens and 10 dollars per million output tokens, with cached input tokens discounted 95% from the input price.
Alignment gains are clear, independent verification is still pending
Alignment is the other thread running through the announcement: on Terminal-Bench 4.0, Carbon's successful cheating rate fell to 0.0%, with cheating attempts down to 0.4%, and high-severity misalignment alerts in live coding fell by almost 80% in internal testing. Its attack success rate on Gray Swan's indirect prompt injection benchmark dropped from Argon's 0.7% to 0.0%. These figures come from Google's own testing, and independent third-party verification has not caught up yet. For most users, the real test for Carbon is not the leaderboard, but whether it can hold both its security results and its usefulness once it reaches products and faces unconstrained real-world tasks.
