Google 宣布推出新一代前沿 AI 模型 Gemini 4 Argon，声称其在复杂工作流、企业知识工作和网络安全防御等领域达到业界领先水平13。该模型支持 100 万 token 输出上限1，输入 token 定价为每百万 2 美元，输出 token 定价为每百万 10 美元1。在多项基准测试中，Argon 表现突出：在软件工程基准测试 DeepSWE v1.1 上得分 77.9%，在 CWE-bench v1 上与第一名并列达到 68%，在 AutomationBench 测试中排名第一、得分 51.3%，在长视频理解基准测试 LVBench 上达到 91.7%12。
目前 Gemini 4 Argon 的访问权限受到严格限制3。Google 首先通过 Fairwind Program 向受信任的网络防御者发布该模型18，随后计划向 API 付费客户和 Google AI Ultra 订阅者提供访问权限1。公司也正在积极参与美国政府的自愿性模型预发布访问流程3。
该模型已在 Google 内部取得显著成果。公司内部数据中心通过 Argon 优化释放了 300 TiB 内存，预计总节省量为 500 TiB 至 1 PiB1。Argon 代理已完成包括 Fuchsia OS Zircon 内核在内的 800,000 多行代码从 C/C++ 到 Rust 的迁移12，libgav1 视频解码器的 Rust 版本性能提升 2.7 倍，达到接近优化 C++ 水平1。此外，网络安全公司 Wiz 利用 Argon 发现了影响全球医疗软件的关键漏洞，而先前的前沿模型未能检测到1。
该模型在经历数月延迟后发布49。发布引发了 Google 内部员工的怀疑态度4610，市场对模型能力和发布策略存在疑虑，导致 Alphabet 股价有所下滑7。
Google has announced Gemini 4 Argon, a new frontier AI model positioned as a major advancement in the company's artificial intelligence capabilities 138. The model is currently available only to a restricted group of "trusted cyber defenders" through Google's Fairwind Program and has not been released to the general public 138. The announcement comes after several months of delays 49.
The model demonstrates substantial performance improvements across multiple domains. In software engineering benchmarks, Gemini 4 Argon achieved 77.9% on DeepSWE v1.1, surpassing competitors including OpenAI's GPT-6 Astra, Anthropic's Fable 5.1, and Opus 5.5 28. It scored 68% on CWE-bench v1, matching the top performer, and ranked first in AutomationBench testing with a score of 51.3% 1. On long-form video understanding, the model reached 91.7% on LVBench 1. The model supports an output context window of one million tokens 15, with pricing set at $2 per million input tokens and $10 per million output tokens 15.
Google has already deployed Gemini 4 Argon internally with significant practical results. The model has optimized internal data centers, freeing 300 terabytes of memory with projected total savings between 500 terabytes and one petabyte 12. It has also completed large-scale code migration projects, including converting over 800,000 lines of code in systems like the Fuchsia OS Zircon kernel from C and C++ to Rust 12. The Rust version of the libgav1 video decoder achieved a 2.7-fold performance improvement, reaching near-optimized C++ levels 1. Additionally, security firm Wiz used Argon to discover critical vulnerabilities in global medical software that previous frontier models failed to detect 1.
Google plans to expand access gradually through a phased rollout strategy, with participation in the U.S. government's voluntary model pre-release program, followed by broader availability to API customers and Google AI Ultra subscribers 1. The model incorporates safety guardrails in response to ongoing artificial intelligence safety discussions 610. The announcement has generated internal skepticism among Google employees, and Alphabet's stock price experienced a decline following reports of internal concerns about the model 7910.
评论
还没有评论，欢迎留下第一条。