Google DeepMind unveiled Gemini 4 Carbon on Wednesday, its strongest model yet, and is holding it back from the general public. The release goes first to a small group of cyber defenders, then expands gradually to other trusted partners under the Fairwind Program. Developers, businesses and consumers have to wait until guardrails get another round of polish.
Safety work runs in parallel. DeepMind says Carbon is already moving through the U.S. government’s voluntary pre-release testing process, and the model was trained to handle complex, high-pressure tasks correctly. It also cuts unsupported claims by roughly half against Argon, the model released two weeks earlier.
On Artificial Analysis Intelligence Index version 5.0, which was recently updated, Carbon takes first place with 59. It lands 88.2% on SEC-bench Pro for locating vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on Collinear’s CWE-bench for root-cause patching. WebDev Arena places it fourth at 1768, up 89 elo points.
Internal results point the same way. Google engineers measured more than twice the secure-code success rate of Argon, and the Chrome team found Carbon produced high-quality patches over 10 times better than another lab’s frontier model. Wiz sees it leading on web penetration testing. Cheating on Terminal-Bench 4.0 fell to 0.0%, with attempts down to 0.4%. High-severity misalignment alerts in live coding dropped nearly 80%. Overreach fell 26% and under-service 29%. On Gray Swan’s indirect prompt injection benchmark, attack success hit 0.0%, against 0.7% for Argon.
List pricing starts where Argon started, at $2 for a million input tokens and $10 for a million output tokens, and cached input carries a 95% discount. Once the introductory window closes, those rates double to $4 and $20.
