Google 发布新模型 Gemini 4 Carbon，在多项基准测试中表现超过 GPT-6 Astra 与 Claude Opus 5.5。该模型在安全基准 SEC-bench Pro 上得分 88.2%、CWE-bench 得分 71.0%，长上下文长距离图遍历基准 GraphWalks 得分 96.2%。
Gemini 4 Carbon, Google's new frontier model, beats GPT-6 Astra and Claude Opus 5.5 on Artificial Analysis.
> Arriving 2 weeks after Gemini 4 Argon, Carbon gains most in security: 88.2% on SEC-bench Pro for finding vulnerabilities and 71.0% on CWE-bench for patching.
54.8% on ExploitGym for exploiting them and 71.0% on CWE-bench for patching them.
Because of that exploit skill, access starts this week with a subset of cyber defenders, then trusted partners through the Fairwind Program, before developers, businesses and consumers get it.
Google says Carbon can work on engineering problems for days, root-causing subtle bugs across large codebases instead of stalling at dead ends.
Cost: introductory $2/$10 per 1M input/output tokens, later doubling to $4 and $20.
On long-context benchmark: on GraphWalks (256K to 1M tokens, BFS F1), Carbon scores 96.2%, against Argon’s 84.2%, GPT-6 Astra’s 71.8% and Claude Opus 5.5’s 66.8%.
来源：Rohan Paul · x.com