Claude Opus 4.6vsOpus 4.5
在 17 个共同 benchmark 中，Claude Opus 4.6 整体领先：Claude Opus 4.6 领先 14 项，Opus 4.5 领先 3 项，持平 0 项，平均分差 +8.99。
Claude Opus 4.6
Anthropic · 2026-02-05 · 推理大模型
Opus 4.5
Anthropic · 2025-11-25 · 推理大模型
Claude Opus 4.614 项(82%)(18%)3 项Opus 4.5
评测分数
按能力类目分组，每组内按分差大小排列；共 17 项。
General Knowledge
Claude Opus 4.6 领先 5/5| 评测项 | Claude Opus 4.6 | Opus 4.5 | 分差 |
|---|---|---|---|
| ARC-AGI-2 | 66.3017 / 62Extended (no tools) | 37.6029 / 62Extended (no tools) | +28.70 |
| ARC-AGI | 9213 / 68Extended (no tools) | 8024 / 68Extended (no tools) | +12 |
| HLE | 5318 / 172Extended (with tools, internet) | 43.2049 / 172Extended (with tools) | +9.80 |
| GPQA Diamond | 91.3115 / 187Extended (no tools) | 8742 / 187Extended (no tools) | +4.31 |
| LiveBench | 76.338 / 115Thinking High (No Tools) | 75.9611 / 115Thinking (No Tools, 64K Budget) | +0.37 |
Agent Level Benchmark
Claude Opus 4.6 领先 2/2| 评测项 | Claude Opus 4.6 | Opus 4.5 | 分差 |
|---|---|---|---|
| τ²-Bench | 91.891 / 43Extended (with tools) | 81.9913 / 43Extended (with tools) | +9.90 |
| τ²-Bench - Telecom | 99.252 / 35Extended (with tools) | 90.7021 / 35Extended (with tools) | +8.55 |
AI Agent - Tool Usage
Claude Opus 4.6 领先 2/2| 评测项 | Claude Opus 4.6 | Opus 4.5 | 分差 |
|---|---|---|---|
| MCP-Atlas | 76.8010 / 27Deep Thinking (With Tools) | 69.8016 / 27Thinking High (With Tools) | +7 |
| Terminal Bench 2.0 | 65.4011 / 47Extended (with tools) | 59.3020 / 47Extended (with tools) | +6.10 |
Coding and Software Engineer
Opus 4.5 领先 2/2| 评测项 | Claude Opus 4.6 | Opus 4.5 | 分差 |
|---|---|---|---|
| LiveCodeBench | 7638 / 123Extended (no tools) | 8712 / 123Extended (with tools) | -11 |
| SWE-bench Verified | 80.8410 / 112Extended (with tools) | 80.909 / 112Extended (with tools) | -0.06 |
Math and Reasoning
Claude Opus 4.6 领先 2/2| 评测项 | Claude Opus 4.6 | Opus 4.5 | 分差 |
|---|---|---|---|
| FrontierMath | 40.707 / 60最高（无工具） | 20.7017 / 60Extended (no tools) | +20 |
| FrontierMath - Tier 4 | 22.9012 / 80最高（无工具） | 4.2040 / 80Normal (No Tools) | +18.70 |
Claw-style Agent Evaluation
Claude Opus 4.6 领先 1/1| 评测项 | Claude Opus 4.6 | Opus 4.5 | 分差 |
|---|---|---|---|
| Pinch Bench | 87.407 / 37Thinking (With Tools) | 87.208 / 37Extended (with tools) | +0.20 |
Instruction Following
Claude Opus 4.6 领先 1/1| 评测项 | Claude Opus 4.6 | Opus 4.5 | 分差 |
|---|---|---|---|
| IF Bench | 941 / 30Extended (no tools) | 5821 / 30Extended (with tools) | +36 |
Multimodal Understanding
Opus 4.5 领先 1/1| 评测项 | Claude Opus 4.6 | Opus 4.5 | 分差 |
|---|---|---|---|
| MMMU | 77.3016 / 29Extended (with tools) | 80.7011 / 29Extended (no tools) | -3.40 |
常识推理
Claude Opus 4.6 领先 1/1| 评测项 | Claude Opus 4.6 | Opus 4.5 | 分差 |
|---|---|---|---|
| Simple Bench | 67.608 / 63Normal (No Tools) | 6212 / 63Extended (no tools) | +5.60 |
规格对比
| 字段 | Claude Opus 4.6 | Opus 4.5 |
|---|---|---|
| 发布机构 | Anthropic | Anthropic |
| 发布时间 | 2026-02-05 | 2025-11-25 |
| 模型类型 | 推理大模型 | 推理大模型 |
| 架构 | 稠密模型 | 稠密模型 |
| 参数规模 | 暂无数据 | 暂无数据 |
| 上下文长度 | 1000K | 200K |
| 最大输出 | 64K | 64K |
API 调用价格
价格优先使用 DataLearner 配置的 API 记录；缺失项不做推测。
| 价格项 | Claude Opus 4.6 | Opus 4.5 |
|---|---|---|
| 文本输入 | $0.5 / 1M tokens | $5 / 1M tokens |
| 文本输出 | $25 / 1M tokens | $25 / 1M tokens |
| 缓存读取 | $0.5 / 1M tokens | $0.5 / 1M tokens |
| 缓存写入 | $10 / 1M tokens | $6.25 / 1M tokens |
小结
- Claude Opus 4.6在以下类目领先:General Knowledge (5/5)、Agent Level Benchmark (2/2)、AI Agent - Tool Usage (2/2)、Math and Reasoning (2/2)、Claw-style Agent Evaluation (1/1)、Instruction Following (1/1)、常识推理 (1/1)
- Opus 4.5在以下类目领先:Coding and Software Engineer (2/2)、Multimodal Understanding (1/1)
17 个共同 benchmark 上，Claude Opus 4.6 平均高出 8.99 分。
单项差距最大的 benchmark：IF Bench — Claude Opus 4.6 94，Opus 4.5 58（分差 +36）。
本页正文由结构化模型、价格与 benchmark 数据生成，不使用实时 LLM 撰写。