GPT-5.5 · HealthBench(48,000+ rubric items across 5,000 multi-turn conversations) · 附件证据(gpt55-e-table7--healthbench)

GPT-5.5 的 HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations 官方出处。见《GPT-5.5 System Card PDF》PDF p.12, Section 5.1, Table 7。证据 gpt55-e-table7--healthbench。

原文与定位

GPT-5.5 · HealthBench(48,000+ rubric items across 5,000 multi-turn conversations),GPT-5.5 model-level safety evaluation配置,长度校正分(0–100) 56.5 score(具体成绩公布时间未确认)。见《GPT-5.5 System Card PDF》PDF p.12, Section 5.1, Table 7。 查看成绩条件

GPT-5.5 · HealthBench(48,000+ rubric items across 5,000 multi-turn conversations),GPT-5.5 model-level safety evaluation配置,未校正分(0–100) 58.4 score(具体成绩公布时间未确认)。见《GPT-5.5 System Card PDF》PDF p.12, Section 5.1, Table 7。 查看成绩条件

GPT-5.5 · HealthBench(48,000+ rubric items across 5,000 multi-turn conversations),GPT-5.5 model-level safety evaluation配置,Mean response length 2313 characters(具体成绩公布时间未确认)。见《GPT-5.5 System Card PDF》PDF p.12, Section 5.1, Table 7。 查看成绩条件

GPT-5.5 · HealthBench(48,000+ rubric items across 5,000 multi-turn conversations),GPT-5.5 model-level safety evaluation配置,相对 GPT-5.4 的长度校正分差 2.5 points(具体成绩公布时间未确认)。见《GPT-5.5 System Card PDF》PDF p.12, Section 5.1, Table 7。 查看成绩条件

GPT-5.5 HealthBench length-adjusted scores, with unadjusted score and mean response length in parentheses: HealthBench 56.5 (58.4, 2313); HealthBench Hard 31.5 (33.8, 2289); HealthBench Consensus 95.6 (95.7, 2259); HealthBench Professional 51.8% (57.2%, 3818). The text says the adjusted changes vs GPT-5.4 are +2.5, +2.4, -0.7, and +3.7 respectively, and separately says the GPT-5.5 Professional mean response length was 3893 characters, conflicting with Table 7’s 3818.
官方材料
GPT-5.5 System Card PDF · PDF 第 12 页
定位
附件 · PDF p.12, Section 5.1, Table 7
材料版本
原文标注日期 2026-04-23;当前材料版本的公布时间未确认。46-page PDF dated April 23, 2026; includes April 24 API-safeguard update notice and August 19, 2026 protein-binding correction.
版本公布
未确认
披露时间
未确认 · 首发/后补未确认
采集时间
历史依据
暂无可靠历史版本依据
引用文本

GPT-5.5 · HealthBench(48,000+ rubric items across 5,000 multi-turn conversations),GPT-5.5 model-level safety evaluation配置,长度校正分(0–100) 56.5 score(具体成绩公布时间未确认);GPT-5.5 · HealthBench(48,000+ rubric items across 5,000 multi-turn conversations),GPT-5.5 model-level safety evaluation配置,未校正分(0–100) 58.4 score(具体成绩公布时间未确认);GPT-5.5 · HealthBench(48,000+ rubric items across 5,000 multi-turn conversations),GPT-5.5 model-level safety evaluation配置,Mean response length 2313 characters(具体成绩公布时间未确认);GPT-5.5 · HealthBench(48,000+ rubric items across 5,000 multi-turn conversations),GPT-5.5 model-level safety evaluation配置,相对 GPT-5.4 的长度校正分差 2.5 points(具体成绩公布时间未确认)。官方出处 GPT-5.5 System Card PDF(附件,原文标注日期 2026-04-23;当前材料版本的公布时间未确认。46-page PDF dated April 23, 2026; includes April 24 API-safeguard update notice and August 19, 2026 protein-binding correction.)PDF p.12, Section 5.1, Table 7,https://deploymentsafety.openai.com/gpt-5-5/gpt-5-5.pdf#page=12。Benchlog https://benchlog.lekshome.cc/evidence/gpt-5-5/gpt55-e-table7--healthbench/

关联成绩与条件

GPT-5.5 · HealthBench 成绩与配置
指标与成绩配置与评测条件证据关系
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。

直接支持该成绩

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。

直接支持该成绩

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.

直接支持该成绩

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.

直接支持该成绩