Claude Sonnet 4
anthropic·2 个数据来源·更新于 2026-11-10
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。
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
8 项已收录成绩 · 保留原始单位和测试条件
Extended thinking (up to 64K tokens) with parallel test-time compute (multiple attempts, internal scoring model selection). Nucleus sampling (top_p 0.95). Based on footnotes 4, 5 and blog appendix.
成绩出处Diamond: Extended thinking (up to 64K tokens) with parallel test-time compute (multiple attempts, internal scoring model selection). Based on footnote 5 and blog appendix.
成绩出处Extended thinking (up to 64K tokens). Average over 14 non-English languages. Based on blog appendix and footnote 3.
成绩出处Parallel test-time compute (multiple attempts, internal scoring model selection). No extended thinking. Based on footnote 5 and SWE-bench methodology for high compute.
成绩出处Extended thinking with tool use (up to 64K tokens, prompt addendum, increased max steps). Based on blog appendix and TAU-bench methodology.
成绩出处Extended thinking with tool use (up to 64K tokens, prompt addendum, increased max steps). Based on blog appendix and TAU-bench methodology.
成绩出处Parallel test-time compute (multiple attempts, internal scoring model selection). No extended thinking. Claude Code as agent framework. Based on footnotes 2 and 5.
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。
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。
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。