GPT-5.6 Sol · 官方评测披露

GPT-5.6 Sol 当前收录 24 项测试的官方记录。已完成列明材料的核查。

模型发布 · GPT Sol

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。

  • GPT-5.6: Frontier intelligence that scales with your ambition | OpenAI · 主文 · 官方英文主文全文及可见表格;原始发布日 2026-07-09;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-5.6 Sol 23 个主文评测面板 已按列出的范围检查。 GPT-5.6 Sol 主文逐点评测提示 已按列出的范围检查。 GPT-5.6 Sol ExploitBench 逐点评测提示 已按列出的范围检查。
  • GPT-5.6 System Card · 附件 · Official page title is GPT-5.6 System Card. Current HTML snapshot; publication date of this page version is not stated.
    公布 未确认 · 采集 · 已核查 Official current HTML saved locally; main content and 48 unique image URLs mapped to figures and visually checked.
  • GPT-5.6 System Card · 附件 · Cover title GPT-5.6 System Card, cover date 2026-07-09; PDF metadata title says GPT-5.6 Preview System Card. Current PDF includes change-log entries dated 2026-08-03 and 2026-08-19; publication date of this current PDF version is not stated.
    公布 未确认 · 采集 · 已核查 All 82 pages text-extracted and reviewed; score tables and the three evaluation-inventory tables were rendered, and every figure was visually checked.

配置

  • GPT-5.6 Sol table result · GPT-5.6 Sol; article table does not print the effort label for this row.
    未知项 Effort may be identified by a matching loaded chart; otherwise it remains unknown.
  • GPT-5.6 Sol Ultra, four agents · Ultra; four agents by default per article footnote.
    未知项 Effort label is not printed in table unless a matching graph identifies it.
  • GPT-5.6 Sol none · GPT-5.6 Sol; reasoning effort none. none
  • GPT-5.6 Sol low · GPT-5.6 Sol; reasoning effort low. low
  • GPT-5.6 Sol medium · GPT-5.6 Sol; reasoning effort medium. medium
  • GPT-5.6 Sol high · GPT-5.6 Sol; reasoning effort high. high
  • GPT-5.6 Sol xhigh · GPT-5.6 Sol; reasoning effort xhigh. xhigh
  • GPT-5.6 Sol max · GPT-5.6 Sol; reasoning effort max. max
  • GPT-5.6 Sol, 1 agent, low · GPT-5.6 Sol; reasoning effort low; 1 agent as labeled in the multi-agent chart. low
  • GPT-5.6 Sol, 4 agents, low · GPT-5.6 Sol; reasoning effort low; 4 agents as labeled in the multi-agent chart. low
  • GPT-5.6 Sol, 16 agents, low · GPT-5.6 Sol; reasoning effort low; 16 agents as labeled in the multi-agent chart. low
  • GPT-5.6 Sol, 1 agent, medium · GPT-5.6 Sol; reasoning effort medium; 1 agent as labeled in the multi-agent chart. medium
  • GPT-5.6 Sol, 4 agents, medium · GPT-5.6 Sol; reasoning effort medium; 4 agents as labeled in the multi-agent chart. medium
  • GPT-5.6 Sol, 16 agents, medium · GPT-5.6 Sol; reasoning effort medium; 16 agents as labeled in the multi-agent chart. medium
  • GPT-5.6 Sol, 1 agent, high · GPT-5.6 Sol; reasoning effort high; 1 agent as labeled in the multi-agent chart. high
  • GPT-5.6 Sol, 4 agents, high · GPT-5.6 Sol; reasoning effort high; 4 agents as labeled in the multi-agent chart. high
  • GPT-5.6 Sol, 16 agents, high · GPT-5.6 Sol; reasoning effort high; 16 agents as labeled in the multi-agent chart. high
  • GPT-5.6 Sol, 1 agent, xhigh · GPT-5.6 Sol; reasoning effort xhigh; 1 agent as labeled in the multi-agent chart. xhigh
  • GPT-5.6 Sol, 4 agents, xhigh · GPT-5.6 Sol; reasoning effort xhigh; 4 agents as labeled in the multi-agent chart. xhigh
  • GPT-5.6 Sol, 16 agents, xhigh · GPT-5.6 Sol; reasoning effort xhigh; 16 agents as labeled in the multi-agent chart. xhigh
  • GPT-5.6 Sol, 1 agent, max · GPT-5.6 Sol; reasoning effort max; 1 agent as labeled in the multi-agent chart. max
  • GPT-5.6 Sol, 4 agents, max · GPT-5.6 Sol; reasoning effort max; 4 agents as labeled in the multi-agent chart. max
  • GPT-5.6 Sol, 16 agents, max · GPT-5.6 Sol; reasoning effort max; 16 agents as labeled in the multi-agent chart. max
  • GPT-5.6 Sol Ultra, four agents, Max · Ultra; four agents by default; reasoning effort Max as labeled in the chart. Max
  • GPT-5.6 Sol HealthBench evaluations · HealthBench and HealthBench Professional-family evaluations; table reports length-adjusted scores.
    未知项 Reasoning effort、Grader version for each GPT-5.6 score、System-level safeguards、Exact benchmark data snapshot

直接前代与后续发布

直接前代 GPT-5.5。 官方关系依据

In order to anticipate changes in rates of misaligned behavior in agentic coding traffic from GPT-5.5 Thinking to GPT-5.6 Sol, the card simulates deployment within OpenAI. It states: “We find that GPT-5.6 Sol, more often than its predecessor, can be overly persistent in pursuing user goals.”
5 项披露延续 · 19 项新增 · 26 项本次未见 · 0 项待核查
GPT-5.5 与 GPT-5.6 Sol 的披露变化
测试已核查状态GPT-5.5GPT-5.6 Sol
Agents’ Last Exam · 55 professional fields; exact task version unspecified新增—本次记录
GDPval-AA · v2新增—本次记录
Big Finance Bench · Version and task protocol unspecified新增—本次记录
Artificial Analysis Intelligence Index · v4.1新增—本次记录
Artificial Analysis Coding Agent Index · v1.1新增—本次记录
SWE-Bench Pro · Subset/version unspecified in source新增—本次记录
DeepSWE · v1.1新增—本次记录
Terminal-Bench · 2.1; source metric reported as score percentage新增—本次记录
GeneBench Pro · Version unspecified in source新增—本次记录
LifeSciBench · Version unspecified in source新增—本次记录
HealthBench Professional · 525 physician-authored conversations延续

来源位置 attachment → attachment、main

已记录的评测条件有变化

前代记录本次记录
OSWorld · 2.0; subset and scoring details unspecified新增—本次记录
BrowseComp · 1,266 题;单代理与多代理按协议区分延续

已记录的评测条件有变化

前代记录本次记录
BenchCAD · Task version and scoring details unspecified新增—本次记录
BenchCAD · Task version unspecified; Python tool protocol新增—本次记录
MMMU-Pro · Standard and Vision formats延续

已记录的评测条件有变化

前代记录本次记录
MMMU-Pro · Tool-enabled protocol; exact version unspecified (gpt56-mmmu-pro-with-tools)新增—本次记录
GDP.pdf · 100 professional PDFs and prompts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
FrontierMath · Tier 1–3 v2新增—本次记录
FrontierMath · Tier 4 v2新增—本次记录
Toolathlon · Version, subset, and metric unspecified in the 2026-07-09 article新增—本次记录
ARC-AGI-3 · Version/subset unspecified in source新增—本次记录
HealthBench · 48,000+ rubric items across 5,000 multi-turn conversations延续

已记录的评测条件有变化

前代记录本次记录
Terminal-Bench · 2.0;代理框架按协议区分本次未见前代记录—
GDPval · 44 occupations; exact task/version unspecified本次未见前代记录—
OSWorld-Verified · Verified benchmark; evaluation protocol noted per source本次未见前代记录—
Toolathlon · Version, subset, and metric unspecified in the 2026-04-23 article本次未见前代记录—
FrontierMath · Tier 1–3; version unspecified本次未见前代记录—
FrontierMath · Tier 4; version unspecified本次未见前代记录—
SWE-Bench Pro · Public本次未见前代记录—
FinanceAgent · v1.1本次未见前代记录—
OfficeQA · Pro 133-question subset本次未见前代记录—
MMMU-Pro · Tool-enabled protocol; exact version unspecified (gpt55-mmmu-pro-with-tools)本次未见前代记录—
MCP-Atlas · 所选材料未明确子集本次未见前代记录—
τ²-bench · Telecom section本次未见前代记录—
GeneBench · 2026-04-23 report; 103 evaluations across 10 domains本次未见前代记录—
BixBench · Version/subset unspecified本次未见前代记录—
Humanity's Last Exam · 2,500-question benchmark本次未见前代记录—
Humanity’s Last Exam · Tool-enabled protocol; exact subset/version unspecified本次未见前代记录—
ARC-AGI-1 · Verified; exact subset not specified本次未见前代记录—
ARC-AGI-2 · ARC Prize Verified · semi-private set本次未见前代记录—
Artificial Analysis Intelligence Index · Exact edition unspecified in chart本次未见前代记录—
AA-LCR · Component of the Artificial Analysis Intelligence Index; version unspecified本次未见前代记录—
AA-Omniscience · Component of the Artificial Analysis Intelligence Index; version unspecified本次未见前代记录—
CritPt · Component of the Artificial Analysis Intelligence Index; version unspecified本次未见前代记录—
GDPval-AA · Artificial Analysis leaderboard本次未见前代记录—
IFBench · Component of the Artificial Analysis Intelligence Index; version unspecified本次未见前代记录—
SciCode · Component of the Artificial Analysis Intelligence Index; version unspecified本次未见前代记录—
Artificial Analysis Coding Index · Exact edition unspecified in article本次未见前代记录—

测试记录 · 24