GPT-5.1 · 官方评测披露

GPT-5.1 当前收录 12 项测试的官方记录。已完成列明材料的核查。

模型发布 · GPT 无后缀型号

官方材料与核查范围

GPT-5.1主文、开发者附录、五页GPT-5.1卡、四页敏感对话附录均已核读;共用GPT-5当前60页基础卡全文/全部45个含表图页已完成,另加入已全文核读的27页Codex-Max后续比较来源。当前列明材料全部完成;不同主体及版本分开记录。 核对时间 。

当前资料不足以还原首发。 首发主文与系统卡说明页的历史HTML已核读;直接PDF附件目前的完成记录仅证明当前原件已读,旧封面日期不能证明首发版本。保留已核读HTML和附件历史缺口;不再发起Wayback请求。

  • GPT-5.1 全新上线:更智能、更具对话感的 ChatGPT | OpenAI · 主文 · Official GPT-5.1 release article; launch-day CDX capture 2025-11-12 19:49:02 UTC, status 200 text/html; current-page revision unknown.
    公布 · 采集 · 已核查 首发日快照及主文已核读。
  • 推出为开发者打造的 GPT-5.1 | OpenAI · 主文 · Official GPT-5.1 developer release article; first same-day CDX capture 2025-11-13 22:01:55 UTC, status 200 text/html; appendix table read.
    公布 · 采集 · 已核查 全文与模型评估表已核读;九行能力成绩逐项录入。
  • GPT-5.1 Instant and GPT-5.1 Thinking System Card Addendum | OpenAI · 附件 · Official GPT-5.1 system-card addendum article; launch-day CDX capture 2025-11-12 21:56:03 UTC, status 200 text/html; linked five-page PDF read.
    公布 · 采集 · 已核查 同日官方系统卡说明页及直接链接已核读。
  • GPT-5.1 Instant and GPT-5.1 Thinking System Card Addendum | OpenAI · 附件 · Five-page system-card PDF dated November 12, 2025; full text and all three safety tables read. Cover/publication dates and release-page links do not establish the exact November 12 launch bytes; no verified launch PDF replay is saved for this source.
    公布 · 采集 · 已核查 五页全文、三张安全评测表已核读;不作常规能力成绩。
  • GPT-5 System Card | OpenAI · 附件 · Linked base card; official current August 13, 2025 revision has 60 pages, fully read with all 45 figure/table pages visually inspected. The August 7 archived 59-page card was read separately; it does not prove the exact November 12 attachment version.
    公布 · 采集 · 已核查 复用GPT-5专属核读代理的完整记录:当前规范卡60页全文已读,45个含表图页均渲染并实际查看,另查看p31正文页;历史59页另行核读并分版本保留。这里只确认基础材料范围,不把GPT-5主体分数归给GPT-5.1。清单见2026-11-16-review-gpt-early/PROGRESS.md及最终REPORT。
  • Addendum to GPT-5 System Card: Sensitive Conversations | OpenAI · 附件 · Four-page GPT-5 safety addendum dated October 27, 2025; full text and tables read. Results compare GPT-5 versions, not GPT-5.1. Cover/publication dates and release-page links do not establish the exact November 12 launch bytes; no verified launch PDF replay is saved for this source.
    公布 · 采集 · 已核查 四页全文及表格已核读;其数字比较 GPT-5 旧版本,不归属 GPT-5.1。
  • GPT-5.1-Codex-Max System Card PDF (later GPT-5.1 comparisons) · 附件 · Current 27-page official PDF dated November 18, 2025; full text pp.1-27 and all tables/figures read. Its GPT-5.1 comparison bars are later disclosures, not GPT-5.1 launch evidence; current revision time unknown.
    公布 · 采集 · 已核查 父侧逐页核读27页全文与全部表图;Figures 8/9/12明确GPT-5.1对照62%/34%/12%。OpenAI PRs、Proof Q&A与风险/安全实验不混普通能力。

配置

  • GPT-5.1 (high) · Developer release table labels high reasoning effort; page separately describes adaptive reasoning in ChatGPT Instant/Thinking. high
  • GPT-5.1 none · API reasoning_effort='none'; default for GPT-5.1 API per developer article. none
    未知项 The article describes qualitative advantages for this mode but does not publish a standardized score for the mode.
  • Codex-Max system-card later comparison · GPT-5.1 comparison bars in the GPT-5.1-Codex-Max System Card; per-model effort not independently identified.
    未知项 比较图没有逐模型明确所有采样与effort设置;不根据对照模型或token预算推定档位。

直接前代与后续发布

直接前代 GPT-5。 官方关系依据

OpenAI states it is upgrading the GPT-5 series; GPT-5.1 Instant and GPT-5.1 Thinking are the two introduced ChatGPT variants.
9 项披露延续 · 3 项新增 · 25 项本次未见 · 0 项待核查
GPT-5 与 GPT-5.1 的披露变化
测试已核查状态GPT-5GPT-5.1
SWE-bench Verified · Verified subset (500 problems)新增—本次记录
GPQA Diamond · 198-question Diamond subset延续

已记录的评测条件有变化

前代记录本次记录
AIME · 2025 edition延续

已记录的评测条件有变化

前代记录本次记录
MMMU · Single-attempt延续

已记录的评测条件有变化

前代记录本次记录
τ²-bench · Telecom section延续

已记录的评测条件有变化

前代记录本次记录
FrontierMath · Set/version unspecified in GPT-5 and GPT-5.1 release tables; Python tool enabled.延续

已记录的评测条件有变化

前代记录本次记录
tau2-bench · Airline category延续

已记录的评测条件有变化

前代记录本次记录
τ²-bench · Retail section延续

已记录的评测条件有变化

前代记录本次记录
BrowseComp Long Context · 128k context延续

已记录的评测条件有变化

前代记录本次记录
SWE-Lancer · IC Diamond set延续

来源位置 attachment、main → attachment

已记录的评测条件有变化

前代记录本次记录
PaperBench · 10-paper subset; each paper requires <10GB external data新增—本次记录
MLE-Bench · 30 competition subset selected from tasks with <50GB data and <10h reference duration新增—本次记录
SWE-bench Verified · GPT-5 发布评测固定子集;500 题中排除 23 题,n=477本次未见前代记录—
Aider Polyglot · Diff or diff-fenced edit format本次未见前代记录—
Humanity's Last Exam · 2,500-question benchmark本次未见前代记录—
HMMT 2025 · 2025 edition; no-tools protocol from GPT-5 developer release本次未见前代记录—
MMMU-Pro · Standard and Vision formats本次未见前代记录—
CharXiv Reasoning · Version/split not stated in selected source本次未见前代记录—
Video-MMMU · Test subset本次未见前代记录—
ERQA · Version unspecified in GPT-5 developer release本次未见前代记录—
Scale MultiChallenge · o3-mini grader protocol from GPT-5 developer release本次未见前代记录—
Internal API instruction-following evaluation · Hard subset; exact set/version unspecified本次未见前代记录—
COLLIE · Version unspecified in GPT-5 developer release本次未见前代记录—
OpenAI MRCR · 2 needles; 128k context本次未见前代记录—
OpenAI MRCR · 2 needles; 256k context本次未见前代记录—
GraphWalks · BFS;0–128K 上下文;题数未注明本次未见前代记录—
GraphWalks · parents;0–128K 上下文;题数未注明本次未见前代记录—
BrowseComp Long Context · 256k context本次未见前代记录—
VideoMME · Long subset, with subtitle category本次未见前代记录—
LongFact-Concepts · GPT-5 developer release; no tools本次未见前代记录—
LongFact-Objects · GPT-5 developer release; no tools本次未见前代记录—
FActScore · GPT-5 developer release; no tools; hallucination-rate metric本次未见前代记录—
SimpleQA · No search-enabled; report F1本次未见前代记录—
HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations本次未见前代记录—
MMLU Language (0-shot) · OpenAI GPT-5 system card, 2025-08-07; MMLU test set translated by professional human translators; chain-of-thought prompting. Section 3.10 says 13 languages, but Table 10 lists 14 language rows.本次未见前代记录—
HealthBench Hard · System-card evaluation; exact dataset version unspecified本次未见前代记录—
HealthBench Consensus · System-card evaluation; exact dataset version unspecified本次未见前代记录—

测试记录 · 12