GPT-5.5 · 官方评测披露

GPT-5.5 当前收录 31 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

列明主文、交互评测图和直接附件已完成模型来源核对;首发历史仍需独立证据。 核对时间 。

当前资料不足以还原首发。 已取得首发当天官方主文回放,但尚未取得并核读全部直接附件的可靠历史正文及外链图像历史字节,无法完整还原首发材料集合;成功回放与失败材料分别保留,不用当前附件替代首发。

  • Introducing GPT-5.5 | OpenAI · 主文 · 官方英文主文全文及可见表格;原始发布日 2026-04-23;当前页面修订时间未知
    公布 未确认 · 采集 · 已核查 英文主文、表格和文字已核读。 已解析图中精确标签;没有从柱高或轴坐标估值。
  • GPT-5.5 System Card | OpenAI · 附件 · System-card entry page; states April 24 API/safeguards update and defines GPT-5.5 Pro as parallel test-time compute.
    公布 未确认 · 采集 · 已核查 GPT-5.5 System Card | OpenAI 已按列出的范围检查。 已核查官方入口文章标注的日期及其指向 Deployment Safety Hub 系统卡的直接链接。
  • GeneBench: Assessing AI Agents for Multi-Stage Inference Problems in Genomics and Quantitative Biology · 附件 · 31-page report; 103 evaluations across 10 domains; table includes full-suite pass rates by reasoning effort.
    公布 未确认 · 采集 · 已核查 GeneBench: Assessing AI Agents for Multi-Stage Inference Problems in Genomics and Quantitative Biology 已按列出的范围检查。
  • Off-diagonal Ramsey-number proof · 附件 · Three-page mathematical proof; reviewed and classified as a research result, not a benchmark.
    公布 未确认 · 采集 · 已核查 Off-diagonal Ramsey-number proof 已按列出的范围检查。
  • OpenAI Preparedness Framework v2 · 附件 · Framework v2; 22 pages; last updated 2025-04-15; risk assessment framework, not a model benchmark.
    公布 未确认 · 采集 · 已核查 OpenAI Preparedness Framework v2 已按列出的范围检查。
  • GPT-5.5 System Card · 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。Official deployment safety page; current HTML captured 2026-11-10 and states the card was updated April 24, 2026 for API safeguards.
    公布 未确认 · 采集 · 已核查 Captured complete deployment safety HTML and all nine numbered sections; checked all visible tables, figures, update text, and references against the PDF and actual linked image files. No archived initial-version snapshot is available to identify when each current-page disclosure first appeared; only explicit dated corrections are assigned a later timing.
  • 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.
    公布 未确认 · 采集 · 已核查 All 46 PDF pages; Sections 1-9, Tables 1-17, Figures 1-22, and reference pages 45-46. Each of the 22 actual linked figure images was visually checked. Figure 2 and Figures 7-11 do not label every plotted point; only explicitly printed numeric values and supported qualitative findings were recorded, with no curve or bar-height estimates. Tables 10, 14, and 16 describe evaluation sets without publishing scores. HealthBench Professional mean response length conflicts between Table 7 (3818 characters) and the following paragraph (3893 characters); both disclosures are preserved.

配置

  • GPT-5.5 xhigh · Reasoning effort xhigh; research environment; output may differ slightly from production ChatGPT in some cases. xhigh
    未知项 A subset of protocol details remain benchmark-specific and unreported.
  • GPT-5.5 Pro · Same underlying model with a setting that uses parallel test-time compute.
    未知项 Parallel compute budget and exact Pro harness settings are not stated.
  • GPT-5.5 none · Interactive chart labels reasoning effort none; chart-specific evaluation settings are otherwise not stated. none
  • GPT-5.5 low · Interactive chart labels reasoning effort low; chart-specific evaluation settings are otherwise not stated. low
  • GPT-5.5 medium · Interactive chart labels reasoning effort medium; chart-specific evaluation settings are otherwise not stated. medium
  • GPT-5.5 high · Interactive chart labels reasoning effort high; chart-specific evaluation settings are otherwise not stated. high
  • GPT-5.5 model-level safety evaluation · 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.

直接前代与后续发布

当前收录未确认直接前代。 官方关系依据

The article repeatedly compares GPT-5.5 directly with GPT-5.4, including per-token latency and benchmark results; the article does not explicitly name a direct predecessor.

测试记录 · 31