IT之家 October 29 news, Google has officially released its new-generation frontier large model Gemini 4 Carbon, focusing on industry-leading cybersecurity capabilities and long-horizon task handling capabilities.
Google calls it its strongest model to date, with a focus on strengthening cybersecurity capabilities and the ability to handle complex, long-duration tasks. The model ranks first in the latest Artificial Analysis Intelligence Index v5.0, with a composite score of 59.
Gemini 4 Carbon’s introductory launch pricing is consistent with that of Gemini 4 Argon released two weeks ago, with cached input Tokens eligible for a 95% discount on the input price.
-
Input: $2 per million Tokens (IT之家 note: approximately 13.3 yuan RMB at current exchange rates)
-
Output: $10 per million Tokens (approximately 66.7 yuan RMB at current exchange rates)
After the introductory discount period ends, regular pricing will resume:
-
Input: $4 per million Tokens (approximately 26.7 yuan RMB at current exchange rates)
-
Output: $20 per million Tokens (approximately 133.3 yuan RMB at current exchange rates)
For long-horizon coding tasks, Carbon is purpose-built for highly difficult agent workflows that need to run continuously for multiple days, and is able to autonomously navigate dead ends, complete root-cause localization in large codebases and low-level systems, write vulnerability patches, and handle other complex multi-step engineering tasks.
On the WebDev Arena front-end full-stack development leaderboard, Carbon improved by 89 Elo points over the previous generation, ranking 4th with 1,769 points.
Cybersecurity capability is the core highlight of this upgrade. Carbon reaches industry-leading levels across the vulnerability discovery, verification, and remediation stages of the full defense lifecycle:
Vulnerability discovery: achieves 88.2% on the SEC-bench Pro benchmark, and can autonomously uncover complex vulnerabilities hidden in real software.
Vulnerability exploit verification: achieves 54.8% on the ExploitGym benchmark, and can automatically determine whether a vulnerability can be exploited, helping security personnel prioritize high-risk vulnerabilities.
Root-cause patch remediation: achieves 71.0% on the CWE-bench benchmark, and can directly turn verified vulnerabilities into runnable remediation code.
Internal test data shows that Carbon’s success rate in generating secure code is more than 2 times higher than 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 other labs’ large frontier models.
In terms of safety alignment, Carbon further strengthens Argon’s multi-layered safety system:
-
The cheating success rate on Terminal-Bench 4.0 drops to 0.0%, and the proportion of cheating attempts drops to 0.4%; in the internal real-time coding environment, high-severity misalignment alerts drop by nearly 80%.
-
Unsupported fabricated claims are reduced by about half compared with Argon; it proactively notes methodological limitations, refuses non-compliant operations when evidence is insufficient, and clearly lists missing information.
-
When faced with ambiguous requirements, Carbon first retrieves relevant materials, feeds back discrepancies to the user, and then executes, rather than forcibly guessing a path.
-
The over-execution rate drops by 26%, and the under-service rate drops by 29%; the end-to-end execution reliability of multi-step plans improves significantly, while strictly complying with project scope constraints.
-
The success rate of indirect prompt injection attacks drops to 0.0%, fully resisting all attack attempts covered in this evaluation.
On release cadence, Google says it is cooperating with the U.S. government’s voluntary pre-release model testing process. Starting this week, Carbon will first be opened to a small group of cybersecurity defenders, and later access for trusted partners will be gradually expanded through the Fairwind program. After collecting feedback and iterating and improving protection mechanisms, it will then be opened to developers, enterprises, and ordinary consumers.