Gemini 3 Pro · 官方评测披露

Gemini 3 Pro 当前收录 26 项测试的官方记录。已完成列明材料的核查。

模型发布 · Gemini Pro

官方材料与核查范围

列明现行主文、能力评测、模型卡和FSF报告,以及本轮取得的首发历史主文与两版模型卡,均已核查完成;当前与首发材料版本分开保存。 核对时间 。

可靠首发材料中已记录 25 项测试。 首发当天官方主文、三幅评测图像、3页评测方法、两版9页模型卡及26页FSF报告均已取得历史正文并核读。首发集合使用实际回放的版本;现行10页模型卡仍作为独立当前材料保存。

  • A new era of intelligence with Gemini 3 · 主文 · Article page collected 2026-11-10; article date 2025-11-18
    公布 · 采集 · 已核查 核心能力与补充材料已分别核对。补充范围:Full extracted release article; checked 'Building Gemini 3 responsibly' and direct Model Card link。
  • Gemini 3 Pro - Evaluations Approach, Methodology & Results · 附件 · 3-page evaluation PDF; results as of November 2025
    公布 未确认 · 采集 · 已核查 3页 PDF 方法及结果表已读取;第3页精确值经渲染核对。
  • Gemini 3 Pro - Model Card · 附件 · 10-page PDF; Model Release November 2025; Last Updated May 2026
    公布 未确认 · 采集 · 已核查 核心能力与补充材料已分别核对。补充范围:Safety, human red teaming and Frontier Safety pp.6-10; Table 1 and safety delta table; checked all 10 extracted pages。说明:The card reports that Deep Think-mode FSF results were consistent with the original assessment but provides no separate Deep Think scores; none are assigned to Gemini 3 Pro.
  • Frontier Safety Framework Report - Gemini 3 Pro (November, 2025) v2 · 附件 · 26-page model-specific report; November 2025 version 2
    公布 未确认 · 采集 · 已核查 核心能力与补充材料已分别核对。补充范围:All 26 pages; tables/figures checked throughout; rendered Figures 1-6 and the chart pages 8, 11, 14, 16, 18 and 19。说明:Figure 4 RE-Bench and Figures 5-6 misalignment plots have no exact point labels; recorded text disclosures only, without estimating bar heights.;The preliminary external wet-lab result on PDF p.9 concerns Gemini 2.5 Pro, not Gemini 3 Pro, so it is excluded from Gemini 3 Pro results.;Appendix 2 contains evaluation-awareness examples but no aggregate benchmark score; recorded as disclosures only.
  • A new era of intelligence with Gemini 3 — launch-day article · 主文 · Official article archived 2025-11-18 16:07:15 UTC; includes independently recovered launch-day benchmark images.
    公布 · 采集 · 已核查 首发当天主文全文已核读;能力主表与 Deep Think 图的同日 GIF 回放、Vending-Bench 2 曲线同日 WebP 回放均已实际查看。曲线没有精确点标签,未估数;Deep Think 分数不归入 Pro 默认配置。三份历史图像和回放响应头保存在本轮整合目录。
  • Gemini 3 Pro Model Card — launch-day early version · 附件 · 9-page official model card archived 20251118111103; separate from the 10-page May 2026 revision.
    公布 未确认 · 采集 · 已核查 历史模型卡9/9页全文已核读,与现行10页版逐段对照;能力主表p.5及安全表pp.7、9已视觉检查。早版主表未列HLE搜索代码分数,晚版列45.8%;两版训练数据描述与评测链接有修订,均单独保留。未将2026年5月新增Deep Think说明倒填首发。
  • Gemini 3 Pro Model Card — launch-day late version · 附件 · 9-page official model card archived 20251118215153; separate from the 10-page May 2026 revision.
    公布 未确认 · 采集 · 已核查 历史模型卡9/9页全文已核读,与现行10页版逐段对照;能力主表p.5及安全表pp.7、9已视觉检查。早版主表未列HLE搜索代码分数,晚版列45.8%;两版训练数据描述与评测链接有修订,均单独保留。未将2026年5月新增Deep Think说明倒填首发。

配置

  • Gemini 3 Pro API default · Gemini API model id gemini-3-pro-preview; default sampling settings. Gemini results are pass@1 and single-attempt unless specified; smaller benchmarks average multiple trials.
    未知项 The evaluation table does not specify a thinking budget or inference-time compute setting.
  • ScreenSpot-Pro · extra_high(材料称coming soon) · Capture-screenshot function calling tool; media_resolution=extra_high. The reviewed evaluation report describes this setting as coming soon to the API.
    未知项 该留存材料未证明此设置当时已通过公共API提供。
  • Gemini 3 Pro ScreenSpot-Pro media_resolution=high · Function calling with a capture-screenshot tool; media_resolution=high.
    未知项 The method page says the high setting was coming soon to the API; no other settings are stated.

直接前代与后续发布

直接前代 Gemini 2.5 Pro (GA)。 官方关系依据

Every generation of Gemini has built on the last, enabling you to do more.
6 项披露延续 · 20 项新增 · 29 项本次未见 · 0 项待核查
Gemini 2.5 Pro (GA) 与 Gemini 3 Pro 的披露变化
测试已核查状态Gemini 2.5 Pro (GA)Gemini 3 Pro
LMArena · Gemini 3 release leaderboard score新增—本次记录
Humanity's Last Exam · 2,500-question benchmark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
GPQA Diamond · 198-question Diamond subset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
MathArena Apex · MathArena leaderboard新增—本次记录
MMMU-Pro · Standard and Vision formats新增—本次记录
Video-MMMU · Gemini 3 Pro model evaluation; recommended media_resolution=HIGH and temperature=0新增—本次记录
SimpleQA Verified · Kaggle official leaderboard新增—本次记录
WebDev Arena · Gemini 3 release leaderboard score新增—本次记录
Terminal-Bench · 2.0;代理框架按协议区分新增—本次记录
SWE-bench Verified · Verified subset (500 problems)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
Vending-Bench · 2新增—本次记录
ARC-AGI-2 · ARC Prize Verified · semi-private set新增—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
CharXiv Reasoning · Validation split新增—本次记录
OmniDocBench 1.5 · Version 1.5; average edit distance across text, formula, table and reading order新增—本次记录
LiveCodeBench Pro · Competitive coding problems from Codeforces, ICPC, and IOI新增—本次记录
tau2-bench · Average across Retail, Airline and Telecom categories新增—本次记录
FACTS Benchmark Suite · Held-out internal grounding, parametric, multimodal and search-retrieval benchmarks新增—本次记录
MMMLU · Multilingual Q&A新增—本次记录
Global PIQA · Commonsense reasoning across 100 languages and cultures新增—本次记录
MRCR-V2 · 8-needle; up to 128K average context延续

已记录的评测条件有变化

前代记录本次记录
MRCR-V2 · 8-needle; pointwise at 1M context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新增—本次记录
tau2-bench · Airline category新增—本次记录
τ²-bench · Telecom section新增—本次记录
ScreenSpot-Pro · 1,581 expert-annotated tasks across 23 professional applications and 3 operating systems新增—本次记录
LiveCodeBench · UI 2025-01-01 to 2025-05-01本次未见前代记录—
Aider Polyglot · Diff or diff-fenced edit format本次未见前代记录—
SimpleQA · No search-enabled; report F1本次未见前代记录—
FACTS Grounding · 2024 release本次未见前代记录—
Global MMLU Lite · Lite; 15 human-translated languages本次未见前代记录—
ECLeKTic · Closed-book QA本次未见前代记录—
HiddenMath-Hard · Hard held-out set本次未见前代记录—
MMMU · Single-attempt本次未见前代记录—
Vibe-Eval · Reka image-understanding set本次未见前代记录—
ZeroBench · Average over 4 runs本次未见前代记录—
BetterChartQA · 9 capability buckets本次未见前代记录—
FLEURS · 53-language subset本次未见前代记录—
CoVoST2 · 21 languages to English本次未见前代记录—
ActivityNet-QA · Test subset本次未见前代记录—
EgoTempo · Test subset本次未见前代记录—
Perception Test · Test subset本次未见前代记录—
QVHighlights · Validation subset; [email protected]本次未见前代记录—
Video-MMMU · Test subset本次未见前代记录—
1H-VideoQA · Test subset本次未见前代记录—
LVBench · Test subset本次未见前代记录—
VideoMME · Audio + visual; Long test subset; 0-shot; max 1024 frames at 1 fps本次未见前代记录—
VATEX · Test subset本次未见前代记录—
VATEX-ZH · Validation subset本次未见前代记录—
YouCook2Cap · Validation subset本次未见前代记录—
Minerva · Video; test subset本次未见前代记录—
Neptune · Video; test subset本次未见前代记录—
VideoMME · Full test set; audio + visual + subtitles本次未见前代记录—
LOFT · Hard retrieval; average context length up to 128K; 300 queries本次未见前代记录—
LOFT · Hard retrieval; 1M context, pointwise value本次未见前代记录—

测试记录 · 26