Gemini 2.5 Pro (GA) · 官方评测披露

Gemini 2.5 Pro (GA) 当前收录 35 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、73页技术报告、6月27日模型卡,以及本轮新增首发主文和70页历史报告均已完整核查;首发集合独立保存,不采用后版新增条件。 核对时间 。

可靠首发材料中已记录 35 项测试。 已取得并核读首发当天主文、主表历史图像和直接附带70页技术报告全部正文、图表、安全章节与附录。6月27日模型卡属于后来材料,不纳入首发范围。GA、Preview、Experimental及checkpoint未明内容分别核对,未把安全图表或早期预览演示归给GA。

  • We're expanding our Gemini 2.5 family of models · 主文 · Article page collected 2026-11-10; article date 2025-06-17
    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; confirmed GA/date and checked safety-related mentions。
  • Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities · 附件 · 73-page PDF; embedded results include Gemini 2.5 Pro GA up to version 06-17; PDF metadata creation 2025-10-16 is not treated as publication date
    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/ART/assurance/FSF/external testing and discussion pp.19-39; appendix benchmark details pp.63-66; case-study and appendix evaluations pp.67-73; rendered and inspected chart pages 28, 30, 32 and relevant figures。说明:Figure 8 memorization rates describe Gemini 2.X/2.5 model-family results without isolating Pro; omitted from Pro results.;Appendix Figures 13 and 15 show timelines without value labels; no values estimated.;Appendix 8.2, 8.4 and 8.5 are demonstrations without stable benchmark sets; they were reviewed and retained in coverage only, not benchmark catalog/results.;In pp.36-38 external testing, autonomous-systems and societal evaluations explicitly identify Preview 05-06, while the external cyber and CBRN evaluation passages do not identify a precise checkpoint. These disclosures are not attributed to GA.;Technical-report Figures 10 and 11 name Gemini 2.5 Pro but do not specify the checkpoint; their results remain under a separate unspecified-report configuration, not the GA configuration.
  • Gemini 2.5 Pro Model Card · 附件 · Last updated 2025-06-27; includes GA, Preview 05-06, and Experimental 03-25 columns
    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 21-page extracted model card; safety/FSF pp.7-21; rendered chart pages 11, 12, 14 and 16; compared GA summary with labeled charts。说明:The GA key-skills summary table prints 10/28 medium and 1/12 hard, while the same card's GA chart prints 15/28 and 6/12; both retained with conflict=true.;The GA deceptive-alignment evaluations are explicitly incomplete in the card.;Preview paragraphs that do not identify a precise checkpoint remain separate from GA and Experimental 03-25.
  • We're expanding our Gemini 2.5 family of models — launch-day article · 主文 · Official launch-day article archived 2025-06-17 16:41:36 UTC; benchmark table GIF archived at 16:41:37 UTC.
    公布 · 采集 · 已核查 已实际取得并核读首发当天主文 https://web.archive.org/web/20250617164136id_/https://blog.google/products/gemini/gemini-2-5-model-family-expands/,以及主表历史GIF https://web.archive.org/web/20250617164137id_/https://storage.googleapis.com/gweb-uniblog-publish-prod/original_images/gemini_2-5_benchmarks_margin_light2x_1.gif。已实际查看历史表格,主文直接附带70页技术报告;未把后来6月27日模型卡当首发文件。
  • Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities — launch-day 70-page version · 附件 · 70-page PDF, archived capture from 2025-06-17; cover dated 2025-06-17
    公布 未确认 · 采集 · 已核查 已实际下载首发当天官方PDF回放 https://web.archive.org/web/20250617164147id_/https://storage.googleapis.com/deepmind-media/gemini/gemini_v2_5_report.pdf;Memento-Datetime Tue, 17 Jun 2025 16:41:47 GMT。70/70页全文、安全章节及附录已核读,成绩表和图表已视觉核查;与现行73页版分开保存。

配置

  • Gemini 2.5 Pro stable API · AI Studio API model id gemini-2.5-pro; default sampling settings. Technical report states Gemini scores are pass@1 and single-attempt unless specified. Aider Polyglot averages 3 trials; SWE-bench uses the internal agentic harness; Vibe-Eval uses Gemini as judge.
    未知项 Thinking budget and any non-default settings for individual evaluations are not stated except where noted in each protocol.

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

Today, we’re taking the next step with our 2.5 Pro and Flash models by releasing them as stable and generally available.

测试记录 · 35