谷歌近日正式向公众推出其最新人工智能模型 Gemini 4 Carbon,该模型此前已率先向部分网络安全防御者开放 1。作为谷歌声称的最强模型,Gemini 4 Carbon 旨在处理极限智力任务,具备与 Anthropic 和 OpenAI 顶级系统竞争的能力 1。此次发布的最大亮点在于引入了“智能体蜂群”(agent swarm)功能 1345。尽管定位高端,谷歌强调该模型比上月发布的前沿模型 Argon 进一步拓展了能力边界,而 Argon 仍适合多数常规软件工程工作 1。
在具体性能与应用场景方面,Gemini 4 Carbon 侧重于网络安全、长周期智能体工作流以及复杂知识工作 2。根据独立基准测试结果,该模型在 CursorBench 4.0、Perplexity’s WANDR 和 Zapier’s AutomationBench 中排名第一,在多模态知识代理(MKA)基准中排名第二 2。在安全指标上,它在 Gray Swan IPI 基准测试中实现了 0% 的提示注入攻击成功率 2。在 Project Fortify 扫描中达到 100% 的召回率,且在开源代码库测试中的召回率较 Argon 提升超过 25% 2。价格策略上,Carbon 与 Argon 保持一致,每百万输入 token 收费 2 美元,每百万输出 token 收费 10 美元 2。
目前,Gemini 4 Carbon 已通过 API、Google AI Studio、Gemini for business、Google Antigravity、Gemini app 以及搜索中的 AI Mode 等平台广泛可用 2。对于消费者和企业用户,该模型正在全球范围内逐步推送,特别针对拥有 AI Ultra 权限的用户在 Gemini app 中开放 345。值得注意的是,关于旧模型 Argon 的状态,多方报道称其依然保留可用 345。此外,谷歌还推出了 Fairwind Program 以支持网络安全应用,并将原有的 10,000 名科学家免费使用计划扩展至覆盖 Gemini 4 Carbon 2。
Google has officially released its latest artificial intelligence model, Gemini 4 Carbon, to the general public 12. The model is now widely available around the world for developers, enterprises, scientists, and consumers 235. This release follows a period of approximately two weeks during which Gemini 4 Carbon was accessible only to select cybersecurity defenders 1. A key new capability introduced with this model is the "agent swarm" feature 1345.
The company describes Gemini 4 Carbon as its most powerful model to date, positioning it as a competitor to top-tier systems from Anthropic and OpenAI 1. While designed to handle extreme intellectual tasks, it serves as an evolution of Argon, another frontier model released by Google last month 1. Argon remains suitable for most software engineering tasks and continues to be available alongside Carbon in certain contexts, such as within the Gemini app for AI Ultra users 1345.
Gemini 4 Carbon demonstrates strong performance in various independent benchmarks. It ranks first on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench, while placing second on the Multimodal Knowledge Agents (MKA) benchmark 2. The model excels specifically in cybersecurity, long-horizon agent workflows, complex coding, and scientific reasoning 2. In security tests, Gemini 4 Carbon achieved a 100% recall rate in Project Fortify scans and improved recall rates by more than 25% compared to Argon when tested on open-source codebases 2. Additionally, it recorded a 0% success rate for prompt injection attacks on the Gray Swan IPI benchmark 2.
Access to Gemini 4 Carbon is provided through multiple platforms, including the API, Google AI Studio, Gemini for business, Google Antigravity, the Gemini app, and AI Mode in Search 2. Pricing is identical to that of Gemini 4 Argon, at $2 per million input tokens and $10 per million output tokens 2. To support specific user groups, Google has expanded the Fairwind Program for cybersecurity professionals and extended a free usage initiative for 10,000 scientists to include the new Carbon model 2.
评论
还没有评论,欢迎留下第一条。