趋近智
参数
-
上下文长度
1,000K
模态
Multimodal
架构
Dense
许可证
Proprietary
发布日期
17 Feb 2026
训练数据截止日期
Aug 2025
注意力结构
Multi-Head Attention
隐藏维度大小
4096
层数
-
注意力头
-
键值头
-
激活函数
-
归一化
-
位置嵌入
Absolute Position Embedding
Claude Sonnet 4.6 is a multimodal foundation model engineered for high-performance agentic workflows, complex software engineering, and large-scale document analysis. As a central component of the Claude 4 model family, it utilizes a dense transformer architecture optimized for balancing computational efficiency with high-order reasoning capabilities. The model is specifically designed to function as a versatile workhorse for enterprise automation, supporting advanced tasks such as autonomous navigation of graphical user interfaces and multi-step agentic planning.
Technically, the model introduces several architectural innovations, including a beta 1-million-token context window that enables the processing of extensive codebases and multi-document datasets in a single inference pass. It features a hybrid reasoning framework that supports both adaptive thinking and extended thinking modes, allowing the model to dynamically allocate internal processing tokens for complex problem-solving. Furthermore, the inclusion of context compaction technology facilitates the efficient management of long-running conversations by summarizing historical context as it approaches architectural limits.
Performance is characterized by significant advancements in computer use, where the model demonstrates human-level proficiency in interacting with standard software environments, including web browsers and spreadsheets. It is highly optimized for the software development lifecycle, providing precise instruction following and a reduction in the common pitfalls of overengineering or output latency. The model is deployed via the Anthropic API and major cloud platforms, offering a scalable solution for developers requiring frontier-level intelligence for high-volume production applications.
Anthropic's fourth generation Claude models with advanced reasoning, extended context windows up to 200K tokens, and configurable thinking effort levels. Features improved safety alignment, nuanced understanding, and sophisticated task completion. Includes Opus (most capable), Sonnet (balanced), and Haiku (fast) variants, with thinking modes that enable transparent chain-of-thought reasoning for complex problems.
排名
#13
| 基准 | 分数 | 排名 |
|---|---|---|
Code Generation HumanEval | 0.96 | 🥇 1 |
Graduate-Level QA GPQA | 0.90 | 🥈 2 |
Professional Knowledge MMLU Pro | 0.87 | ⭐ 5 |
Software Engineering (Verified) SWE-bench Verified | 0.80 | 5 |
Mathematics MATH | 0.85 | 8 |
Grade School Math GSM8K | 0.93 | 8 |
General Knowledge MMLU | 0.89 | ⭐ 9 |
Scientific Reasoning ARC-Challenge | 0.58 | 20 |