Claude Opus 4.8

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

28 项已收录成绩 · 保留原始单位和测试条件

BrowseComp厂商自报 · 越高越好
84.3%

Single-agent with web search, web fetch, code execution, adaptive thinking at max effort, 10M-token limit, context compaction at 200k tokens. Multi-agent (orchestrator + blocking subagents) reaches 88.5%.

成绩出处
CharXiv-R厂商自报 · 越高越好
89.9%

CharXiv Reasoning. Adaptive thinking at max effort with Python tools: 89.9%. Without tools: 80.5%. 1,000 validation questions averaged over 5 runs.

成绩出处
CyberGym厂商自报 · 越高越好
78.8%

Targeted vulnerability reproduction (pass@1) over 1,507 tasks without deployed safeguards. With Tier-3 safeguards: 1.0%.

成绩出处
DeepSearchQA厂商自报 · 越高越好
93.1%

F1 score. Web search, web fetch, code execution, max reasoning effort, adaptive thinking, 1M token budget.

成绩出处
DeepSWE 1.1来源汇总 · 越高越好
59%

DeepSWE v1.1 leaderboard, mini-swe-agent harness, max effort; Pass@1 59% ± 2%.

成绩出处
Finance Agent厂商自报 · 越高越好
53.9%

Agentic financial analysis (Finance Agent v2). Evaluated by Vals AI with adaptive thinking at max effort.

成绩出处
Finance Agent v2来源汇总 · 越高越好
53.92%
成绩出处
FrontierCode 1.1来源汇总 · 越高越好
46.5%

FrontierCode 1.1 current leaderboard; mergeability score at max effort: 46.5%.

成绩出处
FrontierSWE来源汇总 · 越高越好
75%

Claude Code

成绩出处
GPQA厂商自报 · 越高越好
93.6%

GPQA Diamond. Averaged over 25 trials.

成绩出处
Graphwalks BFS >128k厂商自报 · 越高越好
68.1%

F1 score on the 1M-token subset, averaged over 5 trials. 256K subset: 85.9%.

成绩出处
Graphwalks parents >128k厂商自报 · 越高越好
83.3%

F1 score on the 1M-token subset, averaged over 5 trials. 256K subset: 99.3%.

成绩出处
HealthBench Professional厂商自报 · 越高越好
55.8%

Length-adjusted score with adaptive thinking at max effort. Claude Sonnet 4.6 as grader, averaged over 5 trials.

成绩出处
Humanity's Last Exam厂商自报 · 越高越好
57.9%

Multidisciplinary reasoning at max effort. With tools (web search, web fetch, code execution): 57.9%. Without tools: 49.8%.

成绩出处
Include厂商自报 · 越高越好
87.6%

Multilingual regional-knowledge benchmark covering 44 languages. Adaptive thinking enabled with structured JSON output.

成绩出处
LifeSci Diamond (2026-11)来源汇总 · 越高越好
0%

Strict pass rate on 55 regulatory-affairs tasks. Plain API via Vercel AI Gateway with a web_search tool (Exa), provider default settings, one answer per task, graded by GPT-6.1 Sol at medium reasoning effort (Arca, 4 Nov 2026 edition).

成绩出处
LiveBench来源汇总 · 越高越好
77.22%

2026-01-08, Thinking xHigh Effort

成绩出处
MCP Atlas厂商自报 · 越高越好
82.2%

Real-world MCP tool use across production-like servers. Evaluated by Scale AI (April 2026 config: 100 tool-call budget). Mean claim coverage of 86.2%.

成绩出处
OfficeQA Pro厂商自报 · 越高越好
66.2%

Internal agentic harness with extracted-text documents and code-execution tools. OfficeQA full set: 77.6%.

成绩出处
OSWorld-Verified厂商自报 · 越高越好
83.4%

Agentic computer use evaluation. 361 tasks at 100 steps, 1080p resolution, adaptive thinking at max effort. Pass@1 averaged over 5 seeds. Anthropic updated harness methodology (zoom-tool fix, 128K max tokens per turn).

成绩出处
ScreenSpot Pro厂商自报 · 越高越好
87.9%

GUI grounding. Adaptive thinking at max effort with Python tools: 87.9%. Without tools: 82.3%. Averaged over 5 runs.

成绩出处
SWE-bench Multilingual厂商自报 · 越高越好
84.4%

300 problems across 9 programming languages.

成绩出处
SWE-Bench Multimodal厂商自报 · 越高越好
38.4%

Visual context (screenshots, design mockups) added to issue descriptions. Internal harness.

成绩出处
SWE-Bench Pro厂商自报 · 越高越好
69.2%

Agentic coding evaluation. Standard harness, adaptive thinking at max effort, averaged over 5 trials.

成绩出处
SWE-Bench Verified厂商自报 · 越高越好
88.6%

Standard harness with thinking blocks included in sampling, averaged over 5 trials.

成绩出处
Terminal-Bench 2.0厂商自报 · 越高越好
74.6%

Terminal-Bench 2.1 with the Terminus-2 public harness on Daytona, high effort, mean reward over 5 attempts per task (89 tasks, 445 trials).

成绩出处
Terminal-Bench 4.0来源汇总 · 越高越好
23.6%

Terminal-Bench 4.0 public leaderboard; resolution rate 23.6% ± 3.6%; Claude Code agent; max effort.

成绩出处
Toolathlon厂商自报 · 越高越好
59.9%

Pass@1 averaged over 3 trials across all 108 tasks. Internal harness with adaptive thinking at max effort. Pass@3: 67.6%.

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
2026-05-28厂商官网
知识截止
2026-01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 4.8$5$25$0.5
Anthropic$5$25$0.5
DeepInfra$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 4.8SHA-256 dc988e0c2a504fbf…2026-11-10T13:53:19Z查看来源
BenchLMClaude Opus 4.82026-11-10查看来源
ArenaClaude Opus 4.8 (High)2026-11-10T14:07:52Z查看来源
ArenaClaude Opus 4.82026-11-10T14:07:52Z查看来源
ArenaClaude Opus 4.8 (High)2026-11-10T14:07:52Z查看来源
ArenaClaude Opus 4.8 (High)2026-11-10T14:07:52Z查看来源
ArenaClaude Opus 4.82026-11-10T14:07:52Z查看来源
ArenaClaude Opus 4.82026-11-10T14:07:52Z查看来源
ArenaClaude Opus 4.8 (High)2026-11-10T14:07:52Z查看来源
ArenaClaude Opus 4.82026-11-10T14:07:52Z查看来源
OpenRouterClaude Opus 4.8SHA-256 63cbc45064d84060…2026-11-10T13:51:49Z查看来源
厂商官网Claude Opus 4.8SHA-256 872c564432234454…2026-11-10T14:06:48Z查看来源
继续浏览模型排行榜