GPT 无后缀型号 · HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations · 历代记录

按发布时间查看 GPT 无后缀型号 的 4 次发布在 HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations 的成绩、配置与披露。

仅为直接前代间的唯一、有效精确值计算差值,要求单位、指标与明确记录的配置和条件一致。

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 · HealthBench 成绩与配置
指标与成绩配置与评测条件出处
Length-adjusted score67.2%原文 67.2% · 单位 %
公布
GPT-5 thinking (system card)
设置
System-card reasoning model; effort is not stated for the cited evaluation.
未知项
Exact reasoning effort is not reported for this system-card evaluation.
工具
工具条件未记录
推理预算
未记录
框架
未记录
协议
HealthBench, realistic health conversations; score as labeled in Figure 6.
条件
System card target column: gpt-5-thinking (reasoning model; exact reasoning effort is not reported). HealthBench realistic health conversations; Figure 6.
在排名中查看此成绩
Length-adjusted score54.3%原文 54.3% · 单位 %
公布
GPT-5 main
设置
Fast, non-reasoning GPT-5 main model in the system card; exact serving settings are not stated.
未知项
Exact serving configuration is not reported.
工具
工具条件未记录
推理预算
未记录
框架
未记录
协议
HealthBench, realistic health conversations; score as labeled in Figure 6.
条件
System card target column: gpt-5-main (fast non-reasoning model; exact serving settings are not reported). HealthBench realistic health conversations; Figure 6.
在排名中查看此成绩

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-5.2 · HealthBench 成绩与配置
指标与成绩配置与评测条件出处
Length-adjusted score47.6066%原文 0.476066 · 单位 %
公布
GPT-5.2 Instant (system card)
设置
System-card HealthBench table names GPT-5.2 Instant; serving configuration is not stated.
未知项
Exact reasoning effort and serving configuration are not reported.
工具
工具条件未记录
推理预算
未记录
框架
未记录
协议
HealthBench, 5,000 potentially multi-turn conversations scored with example-specific rubrics; source decimal fraction converted to percentage.
条件
GPT-5.2 Instant; reasoning effort and serving configuration are not stated.
Length-adjusted score63.3379%原文 0.633379 · 单位 %
公布
GPT-5.2 Thinking (system card)
设置
System-card HealthBench table names GPT-5.2 Thinking; serving configuration is not stated.
未知项
Exact reasoning effort and serving configuration are not reported.
工具
工具条件未记录
推理预算
未记录
框架
未记录
协议
HealthBench, 5,000 potentially multi-turn conversations scored with example-specific rubrics; source decimal fraction converted to percentage.
条件
GPT-5.2 Thinking; reasoning effort and serving configuration are not stated.

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)

与直接前代的记录缺少满足差值计算的同条件依据。

GPT-5.4 · HealthBench 成绩与配置
指标与成绩配置与评测条件出处
Length-adjusted score62.6%原文 62.6% · 单位 %
公布 未确认
GPT-5.4 system-card evaluation · effort unspecified
设置
GPT-5.4 as named in system-card Table 6; reasoning effort and tool settings are not stated.
未知项
Reasoning effort、Tool access, grader details, and run settings
工具
工具条件未记录
推理预算
未记录
框架
未记录
协议
HealthBench; 5,000 realistic health conversations; example-specific rubrics. Table 6 reports one score without a separate raw/adjusted split.
条件
GPT-5.4, as named in the current saved system card; reasoning effort and evaluation run settings are not stated.
Mean response length3311 characters原文 3311 chars · 单位 characters
公布 未确认
GPT-5.4 system-card evaluation · effort unspecified
设置
GPT-5.4 as named in system-card Table 6; reasoning effort and tool settings are not stated.
未知项
Reasoning effort、Tool access, grader details, and run settings
工具
工具条件未记录
推理预算
未记录
框架
未记录
协议
Mean response length on HealthBench, reported in characters.
条件
GPT-5.4, as named in the current saved system card; reasoning effort and evaluation run settings are not stated.

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-5.5 · HealthBench 成绩与配置
指标与成绩配置与评测条件出处
长度校正分(0–100)56.5 score原文 56.5 · 单位 score
公布 未确认
GPT-5.5 model-level safety evaluation
设置
GPT-5.5 model results for the production release; each benchmark has its stated prompt, task, and safeguard conditions.
未知项
The card does not state one universal system prompt, reasoning effort, or sampling configuration across evaluations.
工具
工具条件未记录
推理预算
未记录
框架
未记录
协议
OpenAI 系统卡 HealthBench 计分;2,000 字符不调整,每多 500 字符减 2.99 分,较短回复对应加分;grader 未说明。
条件
Length-adjusted score; updated HealthBench implementation. 系统卡明确采用 0–100 量表;保留原始分数字符串,不乘除数值。
未校正分(0–100)58.4 score原文 58.4 · 单位 score
公布 未确认
GPT-5.5 model-level safety evaluation
设置
GPT-5.5 model results for the production release; each benchmark has its stated prompt, task, and safeguard conditions.
未知项
The card does not state one universal system prompt, reasoning effort, or sampling configuration across evaluations.
工具
工具条件未记录
推理预算
未记录
框架
未记录
协议
OpenAI 系统卡 HealthBench 计分;2,000 字符不调整,每多 500 字符减 2.99 分,较短回复对应加分;grader 未说明。
条件
Unadjusted score. 系统卡明确采用 0–100 量表;保留原始分数字符串,不乘除数值。
Mean response length2313 characters原文 2313 · 单位 characters
公布 未确认
GPT-5.5 model-level safety evaluation
设置
GPT-5.5 model results for the production release; each benchmark has its stated prompt, task, and safeguard conditions.
未知项
The card does not state one universal system prompt, reasoning effort, or sampling configuration across evaluations.
工具
工具条件未记录
推理预算
未记录
框架
未记录
协议
OpenAI 系统卡 HealthBench 计分;2,000 字符不调整,每多 500 字符减 2.99 分,较短回复对应加分;grader 未说明。
条件
Mean response length for HealthBench.
相对 GPT-5.4 的长度校正分差2.5 points原文 +2.5 · 单位 points
公布 未确认
GPT-5.5 model-level safety evaluation
设置
GPT-5.5 model results for the production release; each benchmark has its stated prompt, task, and safeguard conditions.
未知项
The card does not state one universal system prompt, reasoning effort, or sampling configuration across evaluations.
工具
工具条件未记录
推理预算
未记录
框架
未记录
协议
OpenAI 系统卡 HealthBench 计分;2,000 字符不调整,每多 500 字符减 2.99 分,较短回复对应加分;grader 未说明。
条件
Change in length-adjusted HealthBench score versus GPT-5.4.