谷歌发布Gemini嵌入模型，拓展基础层NLP能力

谷歌发布Gemini嵌入模型，拓展基础层NLP能力 GoogleLaunchesGemini EmbeddingModel, ToppingMTEBWith Best-In-Class PerformanceandUltra-Low Pricing 姚书桥Barney Yao 吴叡霖Louis Ng barney.sq.yao@htisec.com 事件： 2025年7月15日，谷歌正式发布其首个文本嵌入模型Gemini-embedding-001，并宣布全面开放API。该模型以68.37的平均分刷新多语言文本嵌入评测基准MTEB（Massive Text Embedding Benchmark）排行榜，显著领先OpenAI的58.93分，成为当前最强嵌入模型。定价方面，谷歌大幅下探至每百万tokens仅0.15美元，远低于业界主流，面向开发者和独立创作者全面开放。 点评： 模型性能大幅领先，确立嵌入领域新标杆。Gemini-embedding-001在MTEB中涵盖的9大类任务（包括语义检索、分类、聚类、STS、重排等）中全面领先，体现了其泛任务、多语言、跨域的强大表示能力。这一成绩不仅刷新技术上限，也为RAG、搜索、推荐等嵌入应用带来显著性能增益。 价格极致下探，加速嵌入能力大众化。谷歌本次采用极具破坏力的定价策略，百万tokens定价仅0.15美元，相比OpenAI嵌入模型便宜数倍。这一策略将显著降低嵌入模型的调用门槛，为中小企业、教育机构、自由职业者等释放强大生产力，是谷歌少见的“平台式让利”行动。 强化Gemini模型矩阵，构建从理解到执行的全栈闭环。嵌入模型是agentic AI、搜索引擎、智能推荐系统的基础能力。此次发布使得Gemini不再仅仅是生成模型，更拥有“理解-匹配-表达”三位一体的表达与认知能力，构建了谷歌在AI工作流中的底层核心竞争力。 谷歌此次发布Gemini嵌入模型，不仅是一次NLP底层能力的迭代，更是其从内容生成向语义理解全栈平台战略的关键一跃。在AI系统逐渐转向多模态+多Agent协同的方向下，嵌入模型作为任务匹配、上下文传递、知识压缩的核心枢纽，其重要性正被重新定义。我们认为： 嵌入层将成为AI工作流中新的价值高地，代表AI基础设施从算力为王转向语义为王；谷歌具备自然语言、搜索、广告推荐等多场景优势，嵌入模型有望快速落地至Docs、Search、Bard、Cloud等多个产品层，构建强大数据闭环；超低定价策略将迫使OpenAI、Cohere、AWS等对手重新评估其embedding API定价结构，或引发嵌入API服务新一轮价格战。 我们认为谷歌Gemini嵌入模型的发布，是其对OpenAI在语言底座层的一次反超尝试。建议关注未来该模型在VertexAI、Gmail、Docs、Search等核心产品的集成节奏，以及对云服务商间NLP能力差异化格局。 风险提示：1）AI需求不及预期；2）地缘政治环境干扰供应链；3）AI数据中心建造放缓 Fig.2模型已支持100多种语言，其嵌入结果可用于多种任务，例如信息检索和分类。 English Summary: Event: On July 15, 2025, Google officially released its first text embedding model, Gemini-embedding-001, and announced full APIavailability. The model achieved an average score of 68.37 on the Massive Text Embedding Benchmark (MTEB), significantlyoutperformingOpenAI’s 58.93, making it the strongest embedding model to date. In terms of pricing, Google aggressively loweredthe cost to just $0.15 per million tokens, far below the industry average, and made the API widely accessible to developers andindependent creators. Commentary: SOTA performance sets a new benchmark in embeddings. Gemini-embedding-001 leads across all nine major task categories on MTEB—including semantic retrieval, classification, clustering,STS, and reranking—demonstrating exceptionalversatility in cross-task, multilingual, and cross-domain understanding. This resultnot only raises the performance ceiling but also offers meaningful gains for embedding applications like RAG, search, andrecommendations. Disruptive pricing acceleratesdemocratization of embedding capabilities.Google’s ultra-low pricing at $0.15 per million tokens represents a major price disruption—several times cheaper than OpenAI’s offerings. This move dramatically lowers the entry barrier for embedding usage, empowering SMBs, educators, and freelancers. Itreflects a rare “platform-level subsidy” that aligns with Google’s long-term ecosystem strategy. Strengthens the Gemini stack and builds a full-stack semantic foundation.Embedding models are foundational to agentic AI, search engines, and recommendation systems. With this release, Gemini evolvesbeyond a generation model into a “understand–match–generate” framework, reinforcing Google’s semantic infrastructure acrossthe entire AI workflow. Strategic Implications:Google’s release of the Gemini embedding model is not just an advancement in NLP infrastructure—it marks a critical leap from content generation to semantic understanding in its full-stack AI platform strategy. As AI systems trend toward multimodal andmulti-agent collaboration, embedding models are becoming central to task matching, context retention, and knowledgecompression. We believe: The embedding layer will become a new value center in the AI workflow, signaling a shift from compute-dominated tosemantic-dominated infrastructure;Google’s advantage in NLP, search, and ad-tech positions Gemini embeddings for rapid integration into Docs, Search,Bard, and Cloud, creating a strong product-level data flywheel;Its aggressive pricing may force competitors like OpenAI, Cohere, and AWS to reevaluate their API pricing models,potentially triggering a new wave of pricing competition in the embedding-as-a-service market. We see this release as Google’s first major counterattack against OpenAI at the language infrastructure layer. Investors shouldwatch closely how fast Gemini-embedding-001 is integrated into core products like Vertex AI, Gmail, Docs, and Search—and howit reshapes the competitive dynamics in enterprise NLP services. Risks:1) AI demand falls short of expectations; 2) geopolitical environment disrupts supply chain; 3) AI data centre constructionslows down 重要信息披露 本研究报告由海通国际分销，海通国际是由海通国际研究有限公司(HTIRL)，Haitong Securities India Private Limited (HSIPL)，Haitong International Japan K.K. (HTIJKK)和海通国际证券有限公司(HTISCL)的证券研究团队所组成的全球品牌，海通国际证券集团(HTISG)各成员分别在其许可的司法管辖区内从事证券活动。 IMPORTANTDISCLOSURES This research report is distributed by Haitong International, a global brand name for the equity research teams of Haitong International Research Limited (“HTIRL”), Haitong Securities India PrivateLimited (“HSIPL”), Haitong International JapanK.K. (“HTIJKK”