GPT-5.3 Instant · HealthBench(48,000+ rubric items across 5,000 multi-turn conversations) 54.1% · 附件证据(ev-gpt53-instant-healthbench)

GPT-5.3 Instant 的 HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations 官方出处。GPT-5.3 Instant · 54.1%。见《GPT-5.3 Instant System Card PDF | OpenAI》Health Performance / HealthBench / Table 3 / HealthBench row, gpt-5.3-instant column。证据 ev-gpt53-instant-healthbench。

原文与定位

GPT-5.3 Instant · HealthBench(48,000+ rubric items across 5,000 multi-turn conversations),GPT-5.3 Instant配置,Length-adjusted score 54.1%(成绩公布 2026-03-02)。见《GPT-5.3 Instant System Card PDF | OpenAI》Health Performance / HealthBench / Table 3 / HealthBench row, gpt-5.3-instant column。 查看成绩条件

HealthBench: gpt-5.3-instant 54.1% (gpt-5.2-instant 55.4%).
官方材料
GPT-5.3 Instant System Card PDF | OpenAI · PDF 第 4 页
定位
附件 · Health Performance / HealthBench / Table 3 / HealthBench row, gpt-5.3-instant column
材料版本
Current 5-page official PDF read in full; a separate 4-page release-day PDF states evaluations reflect the version shipped February 26, 2026.
版本公布
披露时间
· 已有首发依据
采集时间
历史依据
CDX returned a status-200 application/pdf capture at 20260303195852. The if_ replay returned HTTP 200, application/pdf, and Memento-Datetime Tue, 03 Mar 2026 19:58:52 GMT.
引用文本

GPT-5.3 Instant · HealthBench(48,000+ rubric items across 5,000 multi-turn conversations),GPT-5.3 Instant配置,Length-adjusted score 54.1%(成绩公布 2026-03-02)。官方出处 GPT-5.3 Instant System Card PDF | OpenAI(附件,Current 5-page official PDF read in full; a separate 4-page release-day PDF states evaluations reflect the version shipped February 26, 2026.)Health Performance / HealthBench / Table 3 / HealthBench row, gpt-5.3-instant column,https://deploymentsafety.openai.com/gpt-5-3-instant/gpt-5-3-instant.pdf#page=4。Benchlog https://benchlog.lekshome.cc/evidence/gpt-5-3-instant/ev-gpt53-instant-healthbench/

关联成绩与条件

GPT-5.3 Instant · HealthBench 成绩与配置
指标与成绩配置与评测条件证据关系
Length-adjusted score54.1%原文 54.1% · 单位 %
公布
GPT-5.3 Instant
设置
ChatGPT Instant; API model name gpt-5.3-chat-latest.
未知项
No conventional benchmark configuration is published in the article.
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, potentially multi-turn health conversations; example-specific rubrics.
条件
The release-day system card says GPT-5.3 Instant evaluations reflect the version shipped on 2026-02-26.

直接支持该成绩

在排名中查看此成绩