Google正式发布了Gemini 4 Argon大语言模型，声称该模型具备"前沿推理和改进的复杂工作流程性能"1。该模型拥有100万token输出窗口2，在编码、网络安全、科学、金融和法律等多个基准测试中超越GPT-6 Astra、Claude Fable 5.1和Claude Opus 5.5等竞争对手模型1。
Gemini 4 Argon已向Google AI Pro和Ultra订阅用户提供使用权2，并开始向付费用户推出3。此外，学术科学家和网络安全防护者可申请免费使用6个月2。
该模型在多个领域展现出显著能力。在CWE-Bench漏洞修复评测中排名并列第一，得分68%2；在Box的知识工作评测中相比Gemini 3.1 Pro准确度提升14%2；在财务分析基准测试中99%的数据与分析师构建的参考模型匹配2；在医疗图表评测的200多张复杂合成纵向图表上实现最高召回率和ICD-10代码特异性2。该模型可生成完整多文件应用、微服务架构或重构大型代码库2。
Google已将Gemini 4 Argon应用于内部项目中。该模型在YouTube评论系统优化中发现了冗余emoji渲染问题，消除后将服务器负载降低了45%3；在Project Wing无人机物理模拟器中发现了自2018年以来存在的关键代码漏洞3。Google称该模型经过训练可拒绝恶意指令，配备自动错位监控系统3。
在机器人领域，Gemini 4 Argon在具身推理基准(ERQA)上得分87%，超越所有前沿模型4。该模型可理解物理、空间和3D环境问题，支持Boston Dynamics Spot机器人学习新技能、协作机械臂操作、模拟孪生构建和人类视频动作迁移等应用场景4，支持多视图视频、3D几何、接触力学、本体感觉遥测等多模态输入4。开发者可通过Google Interactions API和Google Antigravity平台使用该模型4。
Google has launched Gemini 4 Argon, an advanced AI model featuring what the company describes as "frontier reasoning and improved performance on complex workflows."1 The model is equipped with a 1 million token output window2 and has begun rolling out to paying users through the Gemini app and Antigravity platform, though access varies by subscription tier.3
According to Google, Gemini 4 Argon outperforms competing models including GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5 across benchmarks in coding, cybersecurity, science, finance, and law.1 The model demonstrates substantial capabilities in code generation, mathematical reasoning, and robotic vision.2 In the CWE-Bench vulnerability remediation evaluation, it ranks tied for first with a score of 68%.2 In Box's knowledge work evaluation, it achieved a 14% accuracy improvement over Gemini 3.1 Pro.2 Financial analysis benchmarks show 99% alignment between the model's data and reference models built by analysts.2 On medical chart evaluation across 200+ complex synthetic longitudinal charts, Gemini 4 Argon achieved the highest recall rate and ICD-10 code specificity.2
The model excels at generating complete multi-file applications, microservice architectures, and refactoring large codebases.2 Google has made Gemini 4 Argon available to Google AI Pro and Ultra subscribers,2 while academic scientists and cybersecurity professionals can apply for six months of free access.2 The company has already deployed the model internally, where it identified redundant emoji rendering in YouTube's comment system, reducing server load by 45%, and discovered a critical code vulnerability in Project Wing's drone physics simulator that had existed since 2018.3
For robotics applications, Gemini 4 Argon scores 87% on the Embodied Reasoning Quotient Assessment (ERQA), surpassing all other frontier models.4 The model supports applications across a robot's full lifecycle, including teaching Boston Dynamics Spot robots new skills, operating collaborative robotic arms, building digital twin simulations, transferring human video motion to robots, converting hand-drawn sketches into physical simulations, and enabling autonomous robot learning improvements.4 It accepts multimodal inputs including multi-view video, 3D geometry, contact mechanics, and proprioceptive telemetry.4 Developers can access Gemini 4 Argon immediately through the Google Interactions API and Google Antigravity platform.4
评论
还没有评论，欢迎留下第一条。