正文 · AI 翻译
IT Home reported on October 29 that Google has officially released its next-generation frontier model, Gemini 4 Carbon, emphasizing industry-leading cybersecurity capabilities and the ability to handle long-horizon tasks.
Google calls it its most powerful model yet, with a particular focus on improving cybersecurity capabilities and the ability to handle complex, long-running tasks. The model ranks first in the latest Artificial Analysis Intelligence Index v5.0, with an overall score of 59.
Gemini 4 Carbon’s introductory launch pricing matches that of Gemini 4 Argon, released two weeks ago. Cached input tokens receive a 95% discount on the input price.
Input: $2 per million tokens (IT Home note: approximately 13.3 yuan at the current exchange rate).
Output: $10 per million tokens (approximately 66.7 yuan at the current exchange rate).
After the introductory promotional period ends, standard pricing will resume:
Input: $4 per million tokens (approximately 26.7 yuan at the current exchange rate).
Output: $20 per million tokens (approximately 133.3 yuan at the current exchange rate).
For long-horizon coding tasks, Carbon is designed for difficult agent workflows that need to run continuously for days. It can independently navigate dead ends and complete complex, multistep engineering tasks such as identifying root causes in large codebases and low-level systems and writing vulnerability patches.
On the WebDev Arena frontend and full-stack development leaderboard, Carbon improved by 89 Elo points over its predecessor, ranking 4th with a score of 1,769.
Cybersecurity capabilities are the central highlight of this upgrade. Carbon has reached an industry-leading level across vulnerability discovery, verification, and remediation throughout the defensive lifecycle:
Internal test data shows that Carbon’s success rate in generating secure code is more than 2 times that of Argon. Tests by the Chrome team show that its performance in generating high-quality patches is more than 10 times better than that of large frontier models from other labs.
On safety alignment, Carbon further strengthens Argon’s multilayer safety system:
On Terminal-Bench 4.0, the cheating success rate fell to 0.0%, while the share of cheating attempts fell to 0.4%. In an internal real-time coding environment, high-severity misalignment alerts fell by nearly 80%.
Unsupported fabricated claims decreased by about half compared with Argon. Carbon proactively flags limitations in its methods, refuses unauthorized actions when evidence is insufficient, and explicitly lists missing information.
When faced with ambiguous requirements, Carbon first retrieves relevant information and reports discrepancies to the user before proceeding, rather than forcing a guess about the approach.
The overexecution rate fell by 26%, and the underdelivery rate fell by 29%. The reliability of end-to-end execution of multistep plans improved significantly, while Carbon strictly adheres to project scope constraints.
The success rate of indirect prompt injection attacks fell to 0.0%, with Carbon resisting every attack attempt covered by this evaluation.
Regarding the release schedule, Google said it is participating in the U.S. government’s voluntary pre-release model testing process. Starting this week, it will first make Carbon available to a small group of cybersecurity defenders, then gradually expand access to trusted partners through the Fairwind program. After gathering feedback and iterating on its safeguards, it will make the model available to developers, businesses, and general consumers.