API 接入Gemini 2.5 Pro
数据驻留：未公开 · 隐私：未公开 · SLA：未公开来源：Gemini API and Google AI Studio available regions ↗选择要核对的信息，直接进入对应章节；各项事实均保留版本、条件和来源。
10/14 项有公开值 · 当前版本 Gemini 2.5 Pro · 数据截止 2026-10-22
页面只展示已审核 Claim；待审核数据不会被当作公开事实。只展示已审核事实；未公开、冲突和过期状态不会被补猜。
只判断选型资料是否足够，不评价模型能力高低。
所有必填事实均通过状态、来源、证据与时效校验。最早核验时间：2026-10-22。
release_datelifecycle_statuscapability_tagsdeployment_optionscontext_window_tokensapi_idinput_price_usd_per_mtokoutput_price_usd_per_mtokavailabilityGemini 2.5 Pro 是 Google DeepMind 的 Gemini 2.5 模型，年鉴当前记录 6 个版本，最新公开记录为 Gemini 2.5 Pro。
权重为未披露，代码为未披露，训练数据为未披露，部署方式为API 服务；这些维度分别判断，不能合并成单一“开源/闭源”结论。
当前公开快照的上下文窗口为 1.0M tokens，架构为 未公开；这里只使用通过审核且来源允许公开采集的 Claim。
2 个维度暂无可公开投影的审核记录
补充来源或线索会进入投稿队列，经编辑核验后才会成为公开事实。当前来源记录 6 个。
技术事实、业务价值、企业治理和证据质量分区呈现
厂商、家族、API ID、日期与生命周期
模态、工具、结构化输出与任务能力
Dense/MoE、参数、层数、专家与 Tokenizer
预训练、SFT、偏好优化与安全对齐
教师/学生、基础模型、轨迹和限制
权重、代码、数据、许可证、部署与 API
API、托管云、地区、集成、隐私与 SLA
输入输出、缓存、Batch 与部署成本边界
测试集、条件、评测方、日期和可比性
数据、权限、审计、合规、安全与人工兜底
行业、部门、工作流、任务与适配条件
公开成本/收益不足时不计算 ROI
27 条已审核事实 · 全部来自公开来源闸门通过的记录
关系只来自年鉴数据库中已登记的归属字段。
当前最新版本的 approved Claim 投影 · 30 条历史事实正在复核
接口兼容与业务适配均以来源绑定记录为准，不由模型名称或能力标签推断。
“兼容接口/平台”仅代表来源明确登记的接入边界，不等同于全部第三方工具都可直接使用。
能力标签说明已登记能力；“擅长”只引用有公开案例、评测或审核映射支持的实际任务。
合同与法务审查 · Gemini 2.5 Pro Preview
下载、API、客户端和更新内容只展示已登记的来源事实；未登记不作推断。
230 个已登记地区 · 230 条版本级记录
只展示来源明确点名模型或版本的案例
业务场景：合同与法务审查
Harvey's public BigLaw Bench evaluation names Gemini 2.5 Pro Preview and reports an 85.02% score. The described legal workflows include transactional due diligence, transaction structuring, litigation drafting, document review and analysis, with reasoning over inputs consisting of hundreds of pages.
ROI：公开投入、持续成本、收益或测量周期不足，未计算。
边界：Google AI Studio vendor showcase describing Harvey's evaluation; the exact API model ID, dated snapshot, test sample, prompt/configuration, scoring harness, costs, production adoption and independent replication are not disclosed. The score is retained as an evaluation result, not converted into ROI or a production efficiency percentage.
保留发布日期精度、生命周期节点和原始来源，未公开日期不会被推断。
模型版本 · shutdown
模型版本 · preview
模型版本 · shutdown
模型版本 · shutdown
模型版本 · active
Google's deprecations table lists June 17, 2025 as the release date for gemini-2.5-pro.
来源：Gemini deprecations ↗Google lists December 2, 2025 as the earliest possible shutdown date and recommends gemini-3.1-pro-preview.
显示最新核验的 8 条 · 共 27 条已审核事实
Gemini 2.5 Pro model page (re-checked 2026-10-22): Model code gemini-2.5-pro; inputs audio, images, video, text and PDF; output text; input token limit 1,048,576; output token limit 65,536; capabilities include caching, code execution, file search, function calling, search grounding, structured outputs, thinking and URL context; stable version gemini-2.5-pro.
Gemini 2.5 Pro model code is gemini-2.5-pro; inputs are audio, images, video, text, and PDF; output is text; input token limit is 1,048,576 and output token limit is 65,536. Caching, code execution, function calling, and structured outputs are supported; thinking is supported.
6 个版本 · 每项可回到来源与技术解释
查看最近版本差异待审核内容不会直接进入公开事实。
高置信度，已绑定来源与证据。
条件：Diff setting; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published.
来源：Gemini 2.5: Our most intelligent AI model ↗Google's official Gemini 2.5 Pro Experimental (03-25) chart reports: Humanity's Last Exam no tools 18.8%; GPQA diamond single attempt 84.0%; AIME 2025 single attempt 86.7%; AIME 2024 single attempt 92.0%; LiveCodeBench v5 single attempt 70.4%; Aider Polyglot whole 74.0% and diff 68.6%; SWE-Bench Verified 63.8%; SimpleQA 52.9%; MMMU single attempt 81.7%; Vibe-Eval (Reka) 69.4%; MRCR 128k average 94.5% and 1M pointwise 83.1%; Global MMLU (Lite) 89.8%. The chart methodology identifies the model ID as gemini-2.5-pro-exp-03-25, uses default sampling, reports Gemini results as pass@1 unless otherwise indicated, and states that Vibe-Eval uses Gemini as a judge. The announcement text explicitly describes SWE-Bench Verified as using a custom agent setup. Evaluation date is not separately published.
引用文本与接口返回都只包含已审核公开事实
Google DeepMind Gemini 2.5 Pro｜最新版本 Gemini 2.5 Pro｜已审核事实 27 条｜公开来源 6 个｜数据截止 2026-10-22｜AI 年鉴｜https://ainianjian.com/models/model.google.gemini-2-5-pro
公开接口返回该模型的版本、已审核事实 27 条及其来源信息，可直接用于脚本或对比。
https://api.ainianjian.com/api/models/model.google.gemini-2-5-pro ↗curl -s 'https://api.ainianjian.com/api/models/model.google.gemini-2-5-pro'
技术规格使用当前版本已审核且来源允许公开采集的事实；模型级开放信息与跨版本地区记录分别标注范围。请按每项适用范围核对，未登记字段显示「未公开」，不使用估算值填充。
Re-checked 2026-10-22: Google still lists the same 230 countries and territories for Gemini API and Google AI Studio, with no names added or removed since the 2026-08 capture.
Google states that the Gemini API and Google AI Studio are available in listed countries and territories. This initial curated slice records the explicitly present United States, Canada, United Kingdom, Japan, Australia, and India entries for the Gemini API / Google AI Studio channel; it does not assert Vertex AI or third-party cloud availability.
Google lists December 2, 2025 as the earliest possible shutdown date and recommends gemini-3.1-pro-preview.
来源：Gemini deprecations ↗Google lists December 2, 2025 as the earliest possible shutdown date and recommends gemini-3.1-pro-preview.
来源：Gemini deprecations ↗模型版本 · preview
Google's official Gemini API guide explicitly documents video understanding: Gemini models can process video input, answer questions about video content, describe events, and refer to timestamps. This seed records the task capability only; it does not infer video generation, OCR accuracy, ASR quality, or a benchmark result.
Gemini 2.5 Pro model code is gemini-2.5-pro; inputs are audio, images, video, text, and PDF; output is text; input token limit is 1,048,576 and output token limit is 65,536. Caching, code execution, function calling, and structured outputs are supported; thinking is supported.
Gemini 2.5 Pro model page (re-checked 2026-10-22): Model code gemini-2.5-pro; inputs audio, images, video, text and PDF; output text; input token limit 1,048,576; output token limit 65,536; capabilities include caching, code execution, file search, function calling, search grounding, structured outputs, thinking and URL context; stable version gemini-2.5-pro.
Google lists paid-tier prices per 1M tokens. Gemini 3.6 Flash standard input/output are $1.50/$7.50 and batch input/output are $0.75/$3.75; context caching and grounding prices are listed separately. Gemini 2.5 Pro uses prompt-size bands at 200k tokens for standard and batch input/output/caching prices.
Google lists paid-tier prices per 1M tokens. Gemini 3.6 Flash standard input/output are $1.50/$7.50 and batch input/output are $0.75/$3.75; context caching and grounding prices are listed separately. Gemini 2.5 Pro uses prompt-size bands at 200k tokens for standard and batch input/output/caching prices.
Google lists gemini-3.6-flash release date July 21, 2026 with no shutdown date announced; gemini-2.5-pro release date June 17, 2025 with no shutdown date announced. The preview IDs gemini-2.5-pro-preview-03-25, -05-06 and -06-05 have release dates March 3, May 6 and June 5, 2025, shutdown date December 2, 2025, and recommended replacement gemini-3.1-pro-preview. Google notes that listed shutdown dates are earliest possible dates.
高置信度，已绑定来源与证据。
条件：Whole-file setting; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published.
来源：Gemini 2.5: Our most intelligent AI model ↗Google's official Gemini 2.5 Pro Experimental (03-25) chart reports: Humanity's Last Exam no tools 18.8%; GPQA diamond single attempt 84.0%; AIME 2025 single attempt 86.7%; AIME 2024 single attempt 92.0%; LiveCodeBench v5 single attempt 70.4%; Aider Polyglot whole 74.0% and diff 68.6%; SWE-Bench Verified 63.8%; SimpleQA 52.9%; MMMU single attempt 81.7%; Vibe-Eval (Reka) 69.4%; MRCR 128k average 94.5% and 1M pointwise 83.1%; Global MMLU (Lite) 89.8%. The chart methodology identifies the model ID as gemini-2.5-pro-exp-03-25, uses default sampling, reports Gemini results as pass@1 unless otherwise indicated, and states that Vibe-Eval uses Gemini as a judge. The announcement text explicitly describes SWE-Bench Verified as using a custom agent setup. Evaluation date is not separately published.
高置信度，已绑定来源与证据。
条件：Single attempt; pass@1; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published.
来源：Gemini 2.5: Our most intelligent AI model ↗Google's official Gemini 2.5 Pro Experimental (03-25) chart reports: Humanity's Last Exam no tools 18.8%; GPQA diamond single attempt 84.0%; AIME 2025 single attempt 86.7%; AIME 2024 single attempt 92.0%; LiveCodeBench v5 single attempt 70.4%; Aider Polyglot whole 74.0% and diff 68.6%; SWE-Bench Verified 63.8%; SimpleQA 52.9%; MMMU single attempt 81.7%; Vibe-Eval (Reka) 69.4%; MRCR 128k average 94.5% and 1M pointwise 83.1%; Global MMLU (Lite) 89.8%. The chart methodology identifies the model ID as gemini-2.5-pro-exp-03-25, uses default sampling, reports Gemini results as pass@1 unless otherwise indicated, and states that Vibe-Eval uses Gemini as a judge. The announcement text explicitly describes SWE-Bench Verified as using a custom agent setup. Evaluation date is not separately published.
高置信度，已绑定来源与证据。
条件：Single attempt; pass@1; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published.
来源：Gemini 2.5: Our most intelligent AI model ↗Google's official Gemini 2.5 Pro Experimental (03-25) chart reports: Humanity's Last Exam no tools 18.8%; GPQA diamond single attempt 84.0%; AIME 2025 single attempt 86.7%; AIME 2024 single attempt 92.0%; LiveCodeBench v5 single attempt 70.4%; Aider Polyglot whole 74.0% and diff 68.6%; SWE-Bench Verified 63.8%; SimpleQA 52.9%; MMMU single attempt 81.7%; Vibe-Eval (Reka) 69.4%; MRCR 128k average 94.5% and 1M pointwise 83.1%; Global MMLU (Lite) 89.8%. The chart methodology identifies the model ID as gemini-2.5-pro-exp-03-25, uses default sampling, reports Gemini results as pass@1 unless otherwise indicated, and states that Vibe-Eval uses Gemini as a judge. The announcement text explicitly describes SWE-Bench Verified as using a custom agent setup. Evaluation date is not separately published.
高置信度，已绑定来源与证据。
条件：Pass@1; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published.
来源：Gemini 2.5: Our most intelligent AI model ↗Google's official Gemini 2.5 Pro Experimental (03-25) chart reports: Humanity's Last Exam no tools 18.8%; GPQA diamond single attempt 84.0%; AIME 2025 single attempt 86.7%; AIME 2024 single attempt 92.0%; LiveCodeBench v5 single attempt 70.4%; Aider Polyglot whole 74.0% and diff 68.6%; SWE-Bench Verified 63.8%; SimpleQA 52.9%; MMMU single attempt 81.7%; Vibe-Eval (Reka) 69.4%; MRCR 128k average 94.5% and 1M pointwise 83.1%; Global MMLU (Lite) 89.8%. The chart methodology identifies the model ID as gemini-2.5-pro-exp-03-25, uses default sampling, reports Gemini results as pass@1 unless otherwise indicated, and states that Vibe-Eval uses Gemini as a judge. The announcement text explicitly describes SWE-Bench Verified as using a custom agent setup. Evaluation date is not separately published.
高置信度，已绑定来源与证据。
条件：Single attempt; pass@1; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published.
来源：Gemini 2.5: Our most intelligent AI model ↗Google's official Gemini 2.5 Pro Experimental (03-25) chart reports: Humanity's Last Exam no tools 18.8%; GPQA diamond single attempt 84.0%; AIME 2025 single attempt 86.7%; AIME 2024 single attempt 92.0%; LiveCodeBench v5 single attempt 70.4%; Aider Polyglot whole 74.0% and diff 68.6%; SWE-Bench Verified 63.8%; SimpleQA 52.9%; MMMU single attempt 81.7%; Vibe-Eval (Reka) 69.4%; MRCR 128k average 94.5% and 1M pointwise 83.1%; Global MMLU (Lite) 89.8%. The chart methodology identifies the model ID as gemini-2.5-pro-exp-03-25, uses default sampling, reports Gemini results as pass@1 unless otherwise indicated, and states that Vibe-Eval uses Gemini as a judge. The announcement text explicitly describes SWE-Bench Verified as using a custom agent setup. Evaluation date is not separately published.
高置信度，已绑定来源与证据。
条件：No tools; Google chart methodology; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published.
来源：Gemini 2.5: Our most intelligent AI model ↗Google's official Gemini 2.5 Pro Experimental (03-25) chart reports: Humanity's Last Exam no tools 18.8%; GPQA diamond single attempt 84.0%; AIME 2025 single attempt 86.7%; AIME 2024 single attempt 92.0%; LiveCodeBench v5 single attempt 70.4%; Aider Polyglot whole 74.0% and diff 68.6%; SWE-Bench Verified 63.8%; SimpleQA 52.9%; MMMU single attempt 81.7%; Vibe-Eval (Reka) 69.4%; MRCR 128k average 94.5% and 1M pointwise 83.1%; Global MMLU (Lite) 89.8%. The chart methodology identifies the model ID as gemini-2.5-pro-exp-03-25, uses default sampling, reports Gemini results as pass@1 unless otherwise indicated, and states that Vibe-Eval uses Gemini as a judge. The announcement text explicitly describes SWE-Bench Verified as using a custom agent setup. Evaluation date is not separately published.
高置信度，已绑定来源与证据。
条件：Single attempt; pass@1; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published.
来源：Gemini 2.5: Our most intelligent AI model ↗Google's official Gemini 2.5 Pro Experimental (03-25) chart reports: Humanity's Last Exam no tools 18.8%; GPQA diamond single attempt 84.0%; AIME 2025 single attempt 86.7%; AIME 2024 single attempt 92.0%; LiveCodeBench v5 single attempt 70.4%; Aider Polyglot whole 74.0% and diff 68.6%; SWE-Bench Verified 63.8%; SimpleQA 52.9%; MMMU single attempt 81.7%; Vibe-Eval (Reka) 69.4%; MRCR 128k average 94.5% and 1M pointwise 83.1%; Global MMLU (Lite) 89.8%. The chart methodology identifies the model ID as gemini-2.5-pro-exp-03-25, uses default sampling, reports Gemini results as pass@1 unless otherwise indicated, and states that Vibe-Eval uses Gemini as a judge. The announcement text explicitly describes SWE-Bench Verified as using a custom agent setup. Evaluation date is not separately published.
高置信度，已绑定来源与证据。
条件：Single attempt; pass@1; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published.
来源：Gemini 2.5: Our most intelligent AI model ↗Google's official Gemini 2.5 Pro Experimental (03-25) chart reports: Humanity's Last Exam no tools 18.8%; GPQA diamond single attempt 84.0%; AIME 2025 single attempt 86.7%; AIME 2024 single attempt 92.0%; LiveCodeBench v5 single attempt 70.4%; Aider Polyglot whole 74.0% and diff 68.6%; SWE-Bench Verified 63.8%; SimpleQA 52.9%; MMMU single attempt 81.7%; Vibe-Eval (Reka) 69.4%; MRCR 128k average 94.5% and 1M pointwise 83.1%; Global MMLU (Lite) 89.8%. The chart methodology identifies the model ID as gemini-2.5-pro-exp-03-25, uses default sampling, reports Gemini results as pass@1 unless otherwise indicated, and states that Vibe-Eval uses Gemini as a judge. The announcement text explicitly describes SWE-Bench Verified as using a custom agent setup. Evaluation date is not separately published.
高置信度，已绑定来源与证据。
条件：128k cumulative/average score; AI Studio API model ID gemini-2.5-pro-exp-03-25; evaluation date not separately published.
来源：Gemini 2.5: Our most intelligent AI model ↗Google's official Gemini 2.5 Pro Experimental (03-25) chart reports: Humanity's Last Exam no tools 18.8%; GPQA diamond single attempt 84.0%; AIME 2025 single attempt 86.7%; AIME 2024 single attempt 92.0%; LiveCodeBench v5 single attempt 70.4%; Aider Polyglot whole 74.0% and diff 68.6%; SWE-Bench Verified 63.8%; SimpleQA 52.9%; MMMU single attempt 81.7%; Vibe-Eval (Reka) 69.4%; MRCR 128k average 94.5% and 1M pointwise 83.1%; Global MMLU (Lite) 89.8%. The chart methodology identifies the model ID as gemini-2.5-pro-exp-03-25, uses default sampling, reports Gemini results as pass@1 unless otherwise indicated, and states that Vibe-Eval uses Gemini as a judge. The announcement text explicitly describes SWE-Bench Verified as using a custom agent setup. Evaluation date is not separately published.
高置信度，已绑定来源与证据。
条件：1M pointwise score; AI Studio API model ID gemini-2.5-pro-exp-03-25; evaluation date not separately published.
来源：Gemini 2.5: Our most intelligent AI model ↗Google's official Gemini 2.5 Pro Experimental (03-25) chart reports: Humanity's Last Exam no tools 18.8%; GPQA diamond single attempt 84.0%; AIME 2025 single attempt 86.7%; AIME 2024 single attempt 92.0%; LiveCodeBench v5 single attempt 70.4%; Aider Polyglot whole 74.0% and diff 68.6%; SWE-Bench Verified 63.8%; SimpleQA 52.9%; MMMU single attempt 81.7%; Vibe-Eval (Reka) 69.4%; MRCR 128k average 94.5% and 1M pointwise 83.1%; Global MMLU (Lite) 89.8%. The chart methodology identifies the model ID as gemini-2.5-pro-exp-03-25, uses default sampling, reports Gemini results as pass@1 unless otherwise indicated, and states that Vibe-Eval uses Gemini as a judge. The announcement text explicitly describes SWE-Bench Verified as using a custom agent setup. Evaluation date is not separately published.
高置信度，已绑定来源与证据。
条件：Pass@1; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published.
来源：Gemini 2.5: Our most intelligent AI model ↗Google's official Gemini 2.5 Pro Experimental (03-25) chart reports: Humanity's Last Exam no tools 18.8%; GPQA diamond single attempt 84.0%; AIME 2025 single attempt 86.7%; AIME 2024 single attempt 92.0%; LiveCodeBench v5 single attempt 70.4%; Aider Polyglot whole 74.0% and diff 68.6%; SWE-Bench Verified 63.8%; SimpleQA 52.9%; MMMU single attempt 81.7%; Vibe-Eval (Reka) 69.4%; MRCR 128k average 94.5% and 1M pointwise 83.1%; Global MMLU (Lite) 89.8%. The chart methodology identifies the model ID as gemini-2.5-pro-exp-03-25, uses default sampling, reports Gemini results as pass@1 unless otherwise indicated, and states that Vibe-Eval uses Gemini as a judge. The announcement text explicitly describes SWE-Bench Verified as using a custom agent setup. Evaluation date is not separately published.
高置信度，已绑定来源与证据。
条件：Custom agent setup; AI Studio API model ID gemini-2.5-pro-exp-03-25; evaluation date not separately published.
来源：Gemini 2.5: Our most intelligent AI model ↗Google's official Gemini 2.5 Pro Experimental (03-25) chart reports: Humanity's Last Exam no tools 18.8%; GPQA diamond single attempt 84.0%; AIME 2025 single attempt 86.7%; AIME 2024 single attempt 92.0%; LiveCodeBench v5 single attempt 70.4%; Aider Polyglot whole 74.0% and diff 68.6%; SWE-Bench Verified 63.8%; SimpleQA 52.9%; MMMU single attempt 81.7%; Vibe-Eval (Reka) 69.4%; MRCR 128k average 94.5% and 1M pointwise 83.1%; Global MMLU (Lite) 89.8%. The chart methodology identifies the model ID as gemini-2.5-pro-exp-03-25, uses default sampling, reports Gemini results as pass@1 unless otherwise indicated, and states that Vibe-Eval uses Gemini as a judge. The announcement text explicitly describes SWE-Bench Verified as using a custom agent setup. Evaluation date is not separately published.
高置信度，已绑定来源与证据。
条件：Gemini used as a judge; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published.
来源：Gemini 2.5: Our most intelligent AI model ↗Google's official Gemini 2.5 Pro Experimental (03-25) chart reports: Humanity's Last Exam no tools 18.8%; GPQA diamond single attempt 84.0%; AIME 2025 single attempt 86.7%; AIME 2024 single attempt 92.0%; LiveCodeBench v5 single attempt 70.4%; Aider Polyglot whole 74.0% and diff 68.6%; SWE-Bench Verified 63.8%; SimpleQA 52.9%; MMMU single attempt 81.7%; Vibe-Eval (Reka) 69.4%; MRCR 128k average 94.5% and 1M pointwise 83.1%; Global MMLU (Lite) 89.8%. The chart methodology identifies the model ID as gemini-2.5-pro-exp-03-25, uses default sampling, reports Gemini results as pass@1 unless otherwise indicated, and states that Vibe-Eval uses Gemini as a judge. The announcement text explicitly describes SWE-Bench Verified as using a custom agent setup. Evaluation date is not separately published.
部分可比 · 测试集：AIME 2024 · 条件：Single attempt; pass@1; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published. · 评测方：Google DeepMind · 日期：未公开
来源：Gemini 2.5: Our most intelligent AI model ↗部分可比 · 测试集：AIME 2025 · 条件：Single attempt; pass@1; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published. · 评测方：Google DeepMind · 日期：未公开
来源：Gemini 2.5: Our most intelligent AI model ↗部分可比 · 测试集：Aider Polyglot · 条件：Diff setting; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published. · 评测方：Google DeepMind · 日期：未公开
来源：Gemini 2.5: Our most intelligent AI model ↗部分可比 · 测试集：Aider Polyglot · 条件：Whole-file setting; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published. · 评测方：Google DeepMind · 日期：未公开
来源：Gemini 2.5: Our most intelligent AI model ↗部分可比 · 测试集：GPQA diamond · 条件：Single attempt; pass@1; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published. · 评测方：Google DeepMind · 日期：未公开
来源：Gemini 2.5: Our most intelligent AI model ↗部分可比 · 测试集：Global MMLU (Lite) · 条件：Pass@1; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published. · 评测方：Google DeepMind · 日期：未公开
来源：Gemini 2.5: Our most intelligent AI model ↗部分可比 · 测试集：Humanity's Last Exam · 条件：No tools; Google chart methodology; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published. · 评测方：Google DeepMind · 日期：未公开
来源：Gemini 2.5: Our most intelligent AI model ↗部分可比 · 测试集：LiveCodeBench v5 · 条件：Single attempt; pass@1; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published. · 评测方：Google DeepMind · 日期：未公开
来源：Gemini 2.5: Our most intelligent AI model ↗部分可比 · 测试集：MMMU · 条件：Single attempt; pass@1; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published. · 评测方：Google DeepMind · 日期：未公开
来源：Gemini 2.5: Our most intelligent AI model ↗部分可比 · 测试集：MRCR · 条件：128k cumulative/average score; AI Studio API model ID gemini-2.5-pro-exp-03-25; evaluation date not separately published. · 评测方：Google DeepMind · 日期：未公开
来源：Gemini 2.5: Our most intelligent AI model ↗部分可比 · 测试集：MRCR · 条件：1M pointwise score; AI Studio API model ID gemini-2.5-pro-exp-03-25; evaluation date not separately published. · 评测方：Google DeepMind · 日期：未公开
来源：Gemini 2.5: Our most intelligent AI model ↗部分可比 · 测试集：SWE-Bench Verified · 条件：Custom agent setup; AI Studio API model ID gemini-2.5-pro-exp-03-25; evaluation date not separately published. · 评测方：Google DeepMind · 日期：未公开
来源：Gemini 2.5: Our most intelligent AI model ↗部分可比 · 测试集：SimpleQA · 条件：Pass@1; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published. · 评测方：Google DeepMind · 日期：未公开
来源：Gemini 2.5: Our most intelligent AI model ↗部分可比 · 测试集：Vibe-Eval (Reka) · 条件：Gemini used as a judge; AI Studio API model ID gemini-2.5-pro-exp-03-25; default sampling; evaluation date not separately published. · 评测方：Gemini as judge · 日期：未公开
来源：Gemini 2.5: Our most intelligent AI model ↗待审核内容不会直接进入公开事实。
待审核内容不会直接进入公开事实。
高置信度，已绑定来源与证据。
条件：Verified on 2026-10-22: the model page lists Model code gemini-2.5-pro as the stable version.
来源：Gemini 2.5 Pro model details ↗Gemini 2.5 Pro model page (re-checked 2026-10-22): Model code gemini-2.5-pro; inputs audio, images, video, text and PDF; output text; input token limit 1,048,576; output token limit 65,536; capabilities include caching, code execution, file search, function calling, search grounding, structured outputs, thinking and URL context; stable version gemini-2.5-pro.
高置信度，已绑定来源与证据。
条件：Explicit public claim projection from agent; this is a documented capability tag, not a benchmark or quality conclusion.
来源：Gemini 2.5 Pro model details ↗Gemini 2.5 Pro model code is gemini-2.5-pro; inputs are audio, images, video, text, and PDF; output is text; input token limit is 1,048,576 and output token limit is 65,536. Caching, code execution, function calling, and structured outputs are supported; thinking is supported.
高置信度，已绑定来源与证据。
条件：Google's official guide explicitly documents video understanding for Gemini API models. The tag means documented video analysis and question answering; it does not mean video generation or a quality ranking.
来源：Gemini API video understanding ↗Google's official Gemini API guide explicitly documents video understanding: Gemini models can process video input, answer questions about video content, describe events, and refer to timestamps. This seed records the task capability only; it does not infer video generation, OCR accuracy, ASR quality, or a benchmark result.
高置信度，已绑定来源与证据。
来源：Gemini 2.5 Pro model details ↗Gemini 2.5 Pro model code is gemini-2.5-pro; inputs are audio, images, video, text, and PDF; output is text; input token limit is 1,048,576 and output token limit is 65,536. Caching, code execution, function calling, and structured outputs are supported; thinking is supported.
高置信度，已绑定来源与证据。
条件：Verified on 2026-10-22: the model page presents the stable gemini-2.5-pro endpoint and Google AI Studio try-it link; no Vertex AI, cloud-marketplace or self-hosted option is asserted here.
来源：Gemini 2.5 Pro model details ↗Gemini 2.5 Pro model page (re-checked 2026-10-22): Model code gemini-2.5-pro; inputs audio, images, video, text and PDF; output text; input token limit 1,048,576; output token limit 65,536; capabilities include caching, code execution, file search, function calling, search grounding, structured outputs, thinking and URL context; stable version gemini-2.5-pro.
高置信度，已绑定来源与证据。
条件：Standard paid tier; prompts > 200k tokens; free tier is free of charge.
来源：Gemini API pricing ↗Google lists paid-tier prices per 1M tokens. Gemini 3.6 Flash standard input/output are $1.50/$7.50 and batch input/output are $0.75/$3.75; context caching and grounding prices are listed separately. Gemini 2.5 Pro uses prompt-size bands at 200k tokens for standard and batch input/output/caching prices.
高置信度，已绑定来源与证据。
条件：Standard paid tier; prompts <= 200k tokens; free tier is free of charge.
来源：Gemini API pricing ↗Google lists paid-tier prices per 1M tokens. Gemini 3.6 Flash standard input/output are $1.50/$7.50 and batch input/output are $0.75/$3.75; context caching and grounding prices are listed separately. Gemini 2.5 Pro uses prompt-size bands at 200k tokens for standard and batch input/output/caching prices.
高置信度，已绑定来源与证据。
条件：Google deprecations table; stable model.
来源：Gemini deprecations ↗Google lists gemini-3.6-flash release date July 21, 2026 with no shutdown date announced; gemini-2.5-pro release date June 17, 2025 with no shutdown date announced. The preview IDs gemini-2.5-pro-preview-03-25, -05-06 and -06-05 have release dates March 3, May 6 and June 5, 2025, shutdown date December 2, 2025, and recommended replacement gemini-3.1-pro-preview. Google notes that listed shutdown dates are earliest possible dates.
高置信度，已绑定来源与证据。
条件：Standard paid tier; prompts > 200k tokens; includes thinking tokens.
来源：Gemini API pricing ↗Google lists paid-tier prices per 1M tokens. Gemini 3.6 Flash standard input/output are $1.50/$7.50 and batch input/output are $0.75/$3.75; context caching and grounding prices are listed separately. Gemini 2.5 Pro uses prompt-size bands at 200k tokens for standard and batch input/output/caching prices.
高置信度，已绑定来源与证据。
条件：Standard paid tier; prompts <= 200k tokens; includes thinking tokens.
来源：Gemini API pricing ↗Google lists paid-tier prices per 1M tokens. Gemini 3.6 Flash standard input/output are $1.50/$7.50 and batch input/output are $0.75/$3.75; context caching and grounding prices are listed separately. Gemini 2.5 Pro uses prompt-size bands at 200k tokens for standard and batch input/output/caching prices.
高置信度，已绑定来源与证据。
条件：Google deprecations table; stable model.
来源：Gemini deprecations ↗Google lists gemini-3.6-flash release date July 21, 2026 with no shutdown date announced; gemini-2.5-pro release date June 17, 2025 with no shutdown date announced. The preview IDs gemini-2.5-pro-preview-03-25, -05-06 and -06-05 have release dates March 3, May 6 and June 5, 2025, shutdown date December 2, 2025, and recommended replacement gemini-3.1-pro-preview. Google notes that listed shutdown dates are earliest possible dates.
高置信度，已绑定来源与证据。
条件：Harvey BigLaw Bench public evaluation; model tested through its API; full sample, prompt, scoring and dated API snapshot are not published
来源：Harvey: validating Gemini 2.5 Pro Preview's advanced legal reasoning with BigLaw Bench ↗Google AI Studio's public case says Harvey evaluated Gemini 2.5 Pro Preview through its API as part of BigLaw Bench, and reports an 85.02% score, the highest score in the displayed evaluation. The page describes transactional due diligence, transaction structuring, litigation drafting, document review and analysis, and reasoning over inputs consisting of hundreds of pages. It does not publish a dated API model ID, sample size, full prompt/configuration, scoring harness or independent replication.
高置信度，已绑定来源与证据。
条件：Harvey's public BigLaw Bench case study; qualitative capability description, not a production productivity claim
来源：Harvey: validating Gemini 2.5 Pro Preview's advanced legal reasoning with BigLaw Bench ↗Google AI Studio's public case says Harvey evaluated Gemini 2.5 Pro Preview through its API as part of BigLaw Bench, and reports an 85.02% score, the highest score in the displayed evaluation. The page describes transactional due diligence, transaction structuring, litigation drafting, document review and analysis, and reasoning over inputs consisting of hundreds of pages. It does not publish a dated API model ID, sample size, full prompt/configuration, scoring harness or independent replication.
部分可比 · 测试集：BigLaw Bench · 条件：Harvey's public case says the model was tested through its API; sample size, prompts, scoring harness, tool configuration and dated API snapshot are not published. · 评测方：Harvey · 日期：未公开
来源：Harvey: validating Gemini 2.5 Pro Preview's advanced legal reasoning with BigLaw Bench ↗