HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · 测试定义与成绩

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations 收录 19 次模型发布的评测披露或成绩。

测试版本与指标

别名 HealthBench

  • Length-adjusted score · % · 越高越好
  • Raw score · % · 越高越好
  • Mean response length · characters · 排序方向未确认
  • 长度校正分(0–100) · score · 越高越好
  • 未校正分(0–100) · score · 越高越好
  • 相对 GPT-5.4 的长度校正分差 · points · 排序方向未确认
  • Unadjusted score · % · 越高越好

各系列最新模型 · 公布成绩

常规展示范围。排名按指标与评测协议分组,已记录的工具、预算及框架分别展示。条件未说明的组保留公布值供查证。

Length-adjusted score %

工具条件未记录 / max / 框架未记录 · HealthBench length-adjusted score; Artificial Analysis; GPT-4.1 grader; max effort.

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · Length-adjusted score · 工具条件未记录 / max / 框架未记录 · HealthBench length-adjusted score; Artificial Analysis; GPT-4.1 grader; max effort.
模型与配置成绩(%)记录与证据
Claude Fable 5.2Max59.6%

工具条件未记录 / low / 框架未记录 · HealthBench length-adjusted score; Artificial Analysis; GPT-4.1 grader; low effort.

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · Length-adjusted score · 工具条件未记录 / low / 框架未记录 · HealthBench length-adjusted score; Artificial Analysis; GPT-4.1 grader; low effort.
模型与配置成绩(%)记录与证据
Claude Fable 5.2Low60.4%

禁止工具 / low / 框架未记录 · HealthBench length-adjusted score; Claude Opus 4.8 grader; no tools or customized system prompt.

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · Length-adjusted score · 禁止工具 / low / 框架未记录 · HealthBench length-adjusted score; Claude Opus 4.8 grader; no tools or customized system prompt.
模型与配置成绩(%)记录与证据
Claude Haiku 5.5Low59.9%

禁止工具 / medium / 框架未记录 · HealthBench length-adjusted score; Claude Opus 4.8 grader; no tools or customized system prompt.

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · Length-adjusted score · 禁止工具 / medium / 框架未记录 · HealthBench length-adjusted score; Claude Opus 4.8 grader; no tools or customized system prompt.
模型与配置成绩(%)记录与证据
Claude Haiku 5.5Medium60.2%

禁止工具 / high / 框架未记录 · HealthBench length-adjusted score; Claude Opus 4.8 grader; no tools or customized system prompt.

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · Length-adjusted score · 禁止工具 / high / 框架未记录 · HealthBench length-adjusted score; Claude Opus 4.8 grader; no tools or customized system prompt.
模型与配置成绩(%)记录与证据
Claude Haiku 5.5High60.3%

禁止工具 / xhigh / 框架未记录 · HealthBench length-adjusted score; Claude Opus 4.8 grader; no tools or customized system prompt.

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · Length-adjusted score · 禁止工具 / xhigh / 框架未记录 · HealthBench length-adjusted score; Claude Opus 4.8 grader; no tools or customized system prompt.
模型与配置成绩(%)记录与证据
Claude Haiku 5.5Xhigh61.1%

禁止工具 / max / 框架未记录 · HealthBench length-adjusted score; Claude Opus 4.8 grader; no tools or customized system prompt.

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · Length-adjusted score · 禁止工具 / max / 框架未记录 · HealthBench length-adjusted score; Claude Opus 4.8 grader; no tools or customized system prompt.
模型与配置成绩(%)记录与证据
Claude Haiku 5.5Max61.6%

工具条件未记录 / low / 框架未记录 · HealthBench length-adjusted score.

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · Length-adjusted score · 工具条件未记录 / low / 框架未记录 · HealthBench length-adjusted score.
模型与配置成绩(%)记录与证据
Claude Sonnet 5.5Low64.9%

工具条件未记录 / medium / 框架未记录 · HealthBench length-adjusted score.

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · Length-adjusted score · 工具条件未记录 / medium / 框架未记录 · HealthBench length-adjusted score.
模型与配置成绩(%)记录与证据
Claude Sonnet 5.5Medium64.9%

工具条件未记录 / high / 框架未记录 · HealthBench length-adjusted score.

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · Length-adjusted score · 工具条件未记录 / high / 框架未记录 · HealthBench length-adjusted score.
模型与配置成绩(%)记录与证据
Claude Sonnet 5.5High64.9%

工具条件未记录 / xhigh / 框架未记录 · HealthBench length-adjusted score.

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · Length-adjusted score · 工具条件未记录 / xhigh / 框架未记录 · HealthBench length-adjusted score.
模型与配置成绩(%)记录与证据
Claude Sonnet 5.5Xhigh64.7%

禁止工具 / max / 框架未记录 · HealthBench; length-adjusted score; five trials; no tools or customized system prompt.

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · Length-adjusted score · 禁止工具 / max / 框架未记录 · HealthBench; length-adjusted score; five trials; no tools or customized system prompt.
模型与配置成绩(%)记录与证据
Claude Opus 5.5Max60.6%

HealthBench: 5,000 realistic, potentially multi-turn health conversations; example-specific rubrics.

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · Length-adjusted score · HealthBench: 5,000 realistic, potentially multi-turn health conversations; example-specific rubrics.
模型与配置成绩(%)记录与证据
GPT-5.3 Instant未注明54.1%

Raw score %

工具条件未记录 / max / 框架未记录 · HealthBench; Artificial Analysis; GPT-4.1 grader; max effort.

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · Raw score · 工具条件未记录 / max / 框架未记录 · HealthBench; Artificial Analysis; GPT-4.1 grader; max effort.
模型与配置成绩(%)记录与证据
Claude Fable 5.2Max58.8%

工具条件未记录 / low / 框架未记录 · HealthBench; Artificial Analysis; GPT-4.1 grader; low effort.

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · Raw score · 工具条件未记录 / low / 框架未记录 · HealthBench; Artificial Analysis; GPT-4.1 grader; low effort.
模型与配置成绩(%)记录与证据
Claude Fable 5.2Low60.1%

工具条件未记录 / max / 框架未记录 · HealthBench, max effort.

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · Raw score · 工具条件未记录 / max / 框架未记录 · HealthBench, max effort.
模型与配置成绩(%)记录与证据
Claude Sonnet 5.5Max69.4%

禁止工具 / max / 框架未记录 · HealthBench; raw score; five trials; no tools or customized system prompt.

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · Raw score · 禁止工具 / max / 框架未记录 · HealthBench; raw score; five trials; no tools or customized system prompt.
模型与配置成绩(%)记录与证据
Claude Opus 5.5Max68.1%

Mean response length characters

System-card HealthBench table: length-adjusted score (unadjusted score, mean response length in characters).

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · Mean response length · System-card HealthBench table: length-adjusted score (unadjusted score, mean response length in characters).
模型与配置成绩(characters)记录与证据
GPT-6.1 Terra未注明1331 characters

System-card Table 7: mean response length paired with adjusted and unadjusted scores.

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · Mean response length · System-card Table 7: mean response length paired with adjusted and unadjusted scores.
模型与配置成绩(characters)记录与证据
GPT-6 Luna未注明1255 characters

长度校正分(0–100) score

System-card HealthBench table: length-adjusted score (unadjusted score, mean response length in characters).

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · 长度校正分(0–100) · System-card HealthBench table: length-adjusted score (unadjusted score, mean response length in characters).
模型与配置成绩(score)记录与证据
GPT-6.1 Terra未注明55.1 score

System-card Table 7: length-adjusted score (unadjusted score, mean response length in characters).

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · 长度校正分(0–100) · System-card Table 7: length-adjusted score (unadjusted score, mean response length in characters).
模型与配置成绩(score)记录与证据
GPT-6 Luna未注明54.5 score

未校正分(0–100) score

System-card HealthBench table: length-adjusted score (unadjusted score, mean response length in characters).

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · 未校正分(0–100) · System-card HealthBench table: length-adjusted score (unadjusted score, mean response length in characters).
模型与配置成绩(score)记录与证据
GPT-6.1 Terra未注明51.1 score

System-card Table 7: unadjusted score paired with the length-adjusted score.

HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · 未校正分(0–100) · System-card Table 7: unadjusted score paired with the length-adjusted score.
模型与配置成绩(score)记录与证据
GPT-6 Luna未注明50 score

历史记录 · 19 次发布

  • Claude Fable 5.2本次发布有 4 条成绩记录,含 4 条精确值(可含历史记录)
  • GPT-6.1 Terra本次发布有 3 条成绩记录,含 3 条精确值(可含历史记录)
  • Claude Haiku 5.5本次发布有 5 条成绩记录,含 5 条精确值(可含历史记录)
  • GPT-6.1 Sol本次发布有 3 条成绩记录,含 3 条精确值(可含历史记录)
  • Claude Sonnet 5.5本次发布有 5 条成绩记录,含 5 条精确值(可含历史记录)
  • Claude Opus 5.5本次发布有 2 条成绩记录,含 2 条精确值(可含历史记录)
  • GPT-6 Luna本次发布有 3 条成绩记录,含 3 条精确值(可含历史记录)
  • GPT-6 Sol本次发布有 3 条成绩记录,含 3 条精确值(可含历史记录)
  • Claude Fable 5.1本次发布有 2 条成绩记录,含 2 条精确值(可含历史记录)
  • GPT-6 Astra本次发布有 3 条成绩记录,含 3 条精确值(可含历史记录)
  • Claude Opus 5本次发布有 2 条成绩记录,含 2 条精确值(可含历史记录)
  • GPT-5.6 Sol本次发布有 3 条成绩记录,含 3 条精确值(可含历史记录)
  • Claude Sonnet 5本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)
  • GPT-5.5本次发布有 4 条成绩记录,含 4 条精确值(可含历史记录)
  • GPT-5.4本次发布有 2 条成绩记录,含 2 条精确值(可含历史记录)
  • GPT-5.3 Instant本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)
  • GPT-5.2本次发布有 2 条成绩记录,含 2 条精确值(可含历史记录)
  • GPT-5本次发布有 2 条成绩记录,含 2 条精确值(可含历史记录)
  • GPT-5 mini本次发布有 1 条成绩记录,含 1 条精确值(可含历史记录)