谷歌正式推出新一代人工智能模型,包括主打软件工程、知识工作和科学工作流平衡的Gemini 4 Neon,以及面向通用机器人处理能力的Gemini 4 Hydrogen 12。在核心定位上,来源1强调Neon旨在结合前沿智能与速度;而来源2指出该版本具备“Carbon质量与Flash速度”的特性,并已集成至GitHub Copilot中 12。此外,Hydrogen版本被描述为旨在取代此前的3.8 Flash模型,以提供更高效的机器人处理能力 2。
在具体性能指标与技术细节方面,仅来源1提供了详细数据。Gemini 4 Neon拥有100万Token的上下文窗口,其定价策略为每百万输入Token收费3美元(缓存输入折扣后为0.15美元/百万),每百万输出Token收费15美元 1。在高思考模式下,该模型相比前代Carbon模型将每任务成本降低了32%,且在顶级竞品间的成本优势达到50%-80% 1。基准测试显示,其在Artificial Analysis Intelligence Index中得分61排名第3,在Vals Index中排名第2,并在Agent Arena Index可操控性维度排名第2 1。安全性方面,该模型在Gray Swan的IPI基准测试中抵抗了所有攻击(成功率0.0%),并在AutoBaxBench困难层级达到76%的最先进水平 1。
关于可用性与专项功能,两个来源提供的信息存在差异但互补。来源1指出,该模型今日开始在Gemini API、Google AI Studio、Gemini for business及Google Antigravity中逐步推出,并通过Fairwind计划向可信防御者提供针对网络安全的代码漏洞发现、验证和修补能力 1。来源2则补充提到,Gemini 4 Neon已直接上线供AI Pro服务用户访问 2。
Google has officially launched its next-generation artificial intelligence model, Gemini 4 Neon, designed to balance frontier-level intelligence with rapid end-to-end task completion in software engineering, knowledge work, and scientific workflows 1. The model is being gradually rolled out starting today on the Gemini API, Google AI Studio, Gemini for Business, and Google Antigravity 1. Additionally, reports indicate that Gemini 4 Neon is now live for AI Pro users and has been integrated into GitHub Copilot 2.
The launch also introduces Gemini 4 Hydrogen, a variant intended for general-purpose robotics that aims to replace the previous 3.8 Flash model 2. Gemini 4 Neon distinguishes itself by offering "Carbon quality at Flash speeds" 2. This positioning is supported by benchmark results showing the model ranks third on the Artificial Analysis Intelligence Index with a score of 61, second on the Vals Index, and second in controllability on the Agent Arena Index 1.
From an economic perspective, Google highlights significant cost reductions compared to the predecessor Carbon model. In high-thinking mode, the cost per task is reduced by 32%, while operating costs are 50% to 80% lower than those of top competing models 1. Pricing for input tokens is set at $3 per million, with cached inputs discounted by 95% to $0.15 per million; output tokens are priced at $15 per million 1. The model supports a context window of one million tokens 1.
Security features are a primary focus for enterprise and developer use cases. As part of the Fairwind Program, Google is providing trusted defenders with specialized capabilities for code vulnerability discovery, verification, and patching 1. Technical evaluations demonstrate strong defensive metrics, with the model resisting all attacks (0.0% success rate) in Gray Swan’s IPI benchmark and achieving a state-of-the-art 76% success rate on the difficult tier of AutoBaxBench 1.
评论
还没有评论,欢迎留下第一条。