Claude Opus 4.6

Anthropic·5 个数据来源·更新于 2026-11-10

上下文窗口1M厂商官网
最大输出128K厂商官网
输入 / 百万 Token$5厂商官网
输出 / 百万 Token$25厂商官网

评测表现

只在同一来源、同一口径下比较。切换来源可查看不同测试结果。

4 个评测来源

当前配置:各基准设置见下表

采集于 2026-11-10查看原始页面
LLM Stats · 公开基准成绩
暂无综合分数当前所选配置的来源分数

不同来源的分数不能直接相加或横向比较;评测任务成本不是 API Token 单价。

逐项基准成绩

27 项已收录成绩 · 保留原始单位和测试条件

AIME 2025厂商自报 · 越高越好
99.79%

Consensus@64 accuracy (most commonly occurring answer across 64 samples). Run independently by Artificial Analysis.

成绩出处
ARC-AGI v2厂商自报 · 越高越好
68.8%

Max effort, 120k thinking budget.

成绩出处
BrowseComp厂商自报 · 越高越好
84%

Web search, web fetch, programmatic tool calling, context compaction triggered at 50k tokens up to 10M total tokens, max reasoning effort, no thinking enabled. Multi-agent harness increased score to 86.8%.

成绩出处
CharXiv-R厂商自报 · 越高越好
77.4%

CharXiv Reasoning. With adaptive thinking, image cropping tool, and max effort. Scores averaged over 5 runs. Without image cropping tool: 68.7%.

成绩出处
CyberGym厂商自报 · 越高越好
73.8%

Cybersecurity vulnerability reproduction. No thinking, default effort, temperature, and top_p. Given a 'think' tool for interleaved thinking in multi-turn evaluations.

成绩出处
DeepSearchQA厂商自报 · 越高越好
91.3%

F1 score. 900-prompt benchmark for evaluating agents on difficult multi-step information-seeking tasks. Programmatic search tools, context compaction, adaptive thinking, max effort up to 10M total tokens.

成绩出处
FigQA厂商自报 · 越高越好
78.3%

LAB-Bench FigQA. With adaptive thinking, image cropping tool, and max effort. Scores averaged over 5 runs. Without image cropping tool (reasoning only): 58.0%.

成绩出处
Finance Agent厂商自报 · 越高越好
60.7%

Agentic financial analysis evaluation.

成绩出处
FrontierSWE来源汇总 · 越高越好
56%

Claude Code

成绩出处
GPQA厂商自报 · 越高越好
91.3%

GPQA Diamond.

成绩出处
Graphwalks BFS >128k厂商自报 · 越高越好
61.5%

GraphWalks BFS 256K subset of 1M. F1 score with 64k output tokens, 1M context window, average of 5 trials. Max output tokens: 61.1. Full 1M variant: 41.2 (64k output).

成绩出处
Graphwalks parents >128k厂商自报 · 越高越好
95.4%

GraphWalks Parents 256K subset of 1M. F1 score with max output tokens, 1M context window, average of 5 trials. At 64k output tokens: 95.1. Full 1M variant: 72.0 (max), 71.1 (64k).

成绩出处
Humanity's Last Exam厂商自报 · 越高越好
53.1%

With tools: web search, web fetch, code execution, programmatic tool calling, context compaction triggered at 50k tokens up to 3M total tokens, max reasoning effort, adaptive thinking enabled. Domain blocklist used for decontamination. Without tools: 40.0%.

成绩出处
Legal Agent Benchmark来源汇总 · 越高越好
4.2%

All-pass rate (Harvey held-out set)

成绩出处
LiveBench来源汇总 · 越高越好
76.33%

2026-01-08, Thinking High Effort

成绩出处
MCP Atlas厂商自报 · 越高越好
62.7%

Scaled tool use. At high effort, reached industry-leading score of 62.7%. At max effort, score is higher.

成绩出处
MMMLU厂商自报 · 越高越好
91.1%

Multilingual Q&A.

成绩出处
MMMU-Pro厂商自报 · 越高越好
77.3%

Visual reasoning with tools: 77.3%. Without tools: 73.9%.

成绩出处
MRCR v2 (8-needle)厂商自报 · 越高越好
76%

MRCR v2 1M 8-needles variant. Mean Match Ratio with max output tokens, average of 5 trials. Anthropic also reports 78.3 with 64k output tokens and 93.0 for the 256K 8-needle subset.

成绩出处
OpenRCA厂商自报 · 越高越好
34.9%

Software failure diagnosis. 1 point if all generated root-cause elements match ground-truth, 0 points if any mismatch. Run on benchmark author's harness, graded using official methodology.

成绩出处
OSWorld厂商自报 · 越高越好
72.7%

Agentic computer use evaluation.

成绩出处
SWE-bench Multilingual厂商自报 · 越高越好
77.83%

Averaged over 25 trials, each run with adaptive thinking, max effort, default sampling settings (temperature, top_p), and with thinking blocks included in sampling results.

成绩出处
SWE-Bench Verified厂商自报 · 越高越好
80.8%

Averaged over 25 trials. With a prompt modification, score reached 81.42%.

成绩出处
Tau2 Retail厂商自报 · 越高越好
91.9%

Agentic tool use evaluation (τ2-bench Retail).

成绩出处
Tau2 Telecom厂商自报 · 越高越好
99.3%

Agentic tool use evaluation (τ2-bench Telecom).

成绩出处
Terminal-Bench 2.0厂商自报 · 越高越好
65.4%

Terminus-2 harness. 1× guaranteed / 3× ceiling resource allocation, 5–15 samples per task across staggered batches.

成绩出处
Vending-Bench 2厂商自报 · 越高越好
801,759%

Final balance in USD. Simulated vending machine business over a full year of operation. Starting balance $5,000.

成绩出处

技术规格

只展示有可靠来源的资料;优先采用厂商官网与官方模型仓库。

发布日期
2026-02-05厂商官网
知识截止
2025-05厂商官网
上下文窗口
1M Token厂商官网
最大输入
1M TokenLLM Stats
最大输出
128K Token厂商官网
输入模态
文本、图像厂商官网
输出模态
文本厂商官网
推理模式
支持厂商官网
许可证
ProprietaryLLM Stats

价格与用量

美元 / 百万 Token。不同服务商或测试配置的报价分别列出。

模型 API 价格,单位为美元每百万 Token
来源 / 服务商输入输出缓存读取
厂商官网 · Claude Opus 4.6$5$25$0.5
Anthropic$5$25待补充

估算一次用量

按表格第一条报价计算,普通文本、无缓存。

预计费用$7.5
不含工具、图片、推理额外计费或阶梯价格;以服务商实际账单为准。

每个数字,都有出处

同名模型通过开发机构和明确版本关联。预览版、日期版本和不同规模模型分别建档。

厂商官网

厂商官方文档 · 基础规格与标准 API 报价

1 条记录最近采集 2026-11-10

LLM Stats

公开模型页与公开榜单;自报成绩单独标识

1 条记录最近采集 2026-11-10

BenchLM

公开评测数据;保留来源原始口径

1 条记录最近采集 2026-11-10

Arena

Arena 官方公开数据集 · CC BY 4.0

8 条记录最近采集 2026-11-10

OpenRouter

公开模型目录与 API 报价

1 条记录最近采集 2026-11-10
查看全部配置与采集记录
来源 / 配置采集时间原始数据
LLM StatsClaude Opus 4.6SHA-256 6bddf449beabf1c1…2026-11-10T13:53:59Z查看来源
BenchLMClaude Opus 4.62026-11-10查看来源
ArenaClaude Opus 4.6 (High)2026-11-10T14:07:52Z查看来源
ArenaClaude Opus 4.62026-11-10T14:07:52Z查看来源
ArenaClaude Opus 4.6 (High)2026-11-10T14:07:52Z查看来源
ArenaClaude Opus 4.6 (High)2026-11-10T14:07:52Z查看来源
ArenaClaude Opus 4.62026-11-10T14:07:52Z查看来源
ArenaClaude Opus 4.62026-11-10T14:07:52Z查看来源
ArenaClaude Opus 4.6 (High)2026-11-10T14:07:52Z查看来源
ArenaClaude Opus 4.62026-11-10T14:07:52Z查看来源
OpenRouterClaude Opus 4.6SHA-256 63cbc45064d84060…2026-11-10T13:51:49Z查看来源
厂商官网Claude Opus 4.6SHA-256 b4523e3290954239…2026-11-10T14:06:49Z查看来源
继续浏览模型排行榜