[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-gemini-3-7-flash-coding-agent-fast":3,"news-related-16856034-439d-4915-aed4-80b42ae09c68":38},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"16856034-439d-4915-aed4-80b42ae09c68","Gemini 3.7 Flash：FrontierCode 43.6%，价格腰斩","Google 于 2026 年 8 月 13 日发布 Gemini 3.7 Flash,这是 Gemini 3.6 Flash 之后仅 3 周就推出的最新「工作马」模型。在代码、长上下文知识任务、Web 开发基准上带来明显跃迁:FrontierCode 1.1 Main 从 34.4% 提到 43.6%,DeepSWE v1.1 从 49.0% 提到 65.3%,WebDev Arena Elo 从 1538 升至 1588,GDP.pdf 从 22.0% 提到 34.0%。3.7 Flash 同步推出半价促销——每百万输入 token 0.75 美元、输出 3.75 美元,且支持自定义「思考强度」配置。Gemini Spark 在 160 个国家即日起切到 3.7 Flash。","## 三周迭代:Google 把 Flash 的发布节奏卷到了新量级\n\n2026 年 8 月 13 日,Google AI Blog 发布 Gemini 3.7 Flash,这是 Gemini 3.6 Flash(2026 年 7 月 21 日发布)之后仅 3 周就推出的新版本。Google 高级产品总监 Tulsee Doshi 在官方博文中称 3.7 Flash 是「我们迄今为止最智能的工作马模型」,专为编码和 Agent 工作流设计。\n\n把发布节奏压到 3 周,在 Flash 系列里是第一次。3.6 Flash、3.5 Flash Lite、3.5 Flash Cyber 三个模型 7 月 21 日同期上线,3.7 Flash 立刻跟上。Google 给出的解释是「开发者反馈 + 算法创新」,但这种节奏本身就是一个工程信号:前线的反馈循环已经被 AI 改写到以周为单位。\n\n## 代码能力的关键 benchmark 跳变\n\n3.7 Flash 的核心提升集中在代码和长上下文知识任务,以下数据来自 Google 官方博客:[Introducing Gemini 3.7 Flash](https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Fmodels-and-research\u002Fgemini-models\u002Fintroducing-gemini-3-7-flash\u002F)。\n\n- **FrontierCode 1.1 Main**:43.6%(3.7)对比 34.4%(3.6),9.2 个百分点的提升\n- **DeepSWE v1.1**:65.3% 对比 49.0%,提升 16.3 个百分点\n- **WebDev Arena Elo**:1588 对比 1538\n- **GDP.pdf 复杂文档理解**:34.0% 对比 22.0%,提升 12 个百分点\n- **AutomationBench(Zapier 真实工作流)**:30.4% 对比 17.0%,提升 13.4 个百分点\n\n特别值得注意的是 DeepSWE v1.1 和 AutomationBench——前者是 Cognition 推出的软件工程基准,后者是 Zapier 用来衡量「模型能不能在真实业务工作流里代替人」的评估。3.7 Flash 在这两个基准上的跃迁幅度都超过 13 个百分点,说明 Google 不只是把 Flash 的代码能力做「更稳」,而是直接往「能干更复杂的工程」上推。\n\n## 半价促销:把价格再次腰斩\n\n3.7 Flash 的定价策略同样激进。Google 给出「introductory price」——**每百万输入 token 0.75 美元、每百万输出 token 3.75 美元**,有效期到 2026 年 12 月 31 日。**这是 3.6 Flash 的一半价格**。2027 年 1 月 1 日之后,价格回到 1.50 美元 \u002F 7.50 美元每百万 token。\n\n这把「价格减半 + 能力提升」同时打包,实质是 Google 在用价格换 Agent 时代的开发者份额。Box、Browser Use、Cartwheel、Harvey、Hebbia、LangChain、Nunu.ai、Open Code、Pydantic、Stanford 生物学系等一批企业客户和 Agent 框架被官方作为早期客户列出。Stanford 生物系的引用尤其值得关注——它意味着 3.7 Flash 在「知识密集型学科」的可信度被一线研究机构背书。\n\n## 可配置的思考强度\n\n3.7 Flash 的 Model Card 里明确写了一句:**支持可自定义的思考配置,以控制质量、成本和延迟之间的权衡**。\n\n这跟 Gemini 3.5 Pro 那种「全自动 thinking」路线不同——3.7 Flash 把思考强度暴露给开发者,让企业可以根据任务复杂度调节,而不是按 token 计费统一拉到顶配。对于真实生产部署,这意味着低延迟场景(自动补全、UI 渲染)可以关掉深度推理,复杂任务(代码重构、合同审查)再开满,边际成本结构更可预测。\n\n## 160 个国家的 Gemini Spark 即日切换\n\n3.7 Flash 即日起成为 Gemini Spark 的底层模型。Spark 是 Google 在 I\u002FO 2026 上推出的「7×24 个人 AI Agent」,面向 Google AI Pro 和 Ultra 订阅用户,在 160 个国家上线。Spark 这次切到 3.7 Flash,意味着所有 Spark 用户都会「不知不觉」地用上更强的工具调用和工作流编排能力——这是 Google 把模型升级与产品升级同步推进的一次操作。\n\n## 安全侧:CBRN + 网络攻击双加固\n\n3.7 Flash 同步更新了 Frontier Safety 的安全边界,**对化学、生物、放射、核(CBRN)和网络攻击滥用场景做了新的针对性防护**。Google 同时强调这一加固是「在阻止滥用的同时保留有益用途」,并提供了生物安全和网络项目的公开文档链接。安全框架跟产品同步迭代是过去 18 个月里前沿模型的标准做法,但 3.7 Flash 把这一步骤写在博客头条,说明 Google 是在回应监管和投资人最近对模型滥用的持续关注。\n\n## 接入入口\n\n开发者可通过 Google AI Studio 的 Gemini API、Android Studio、Antigravity 平台调用 3.7 Flash;企业可通过 Gemini Enterprise Agent Platform 和 Gemini Enterprise app 接入;个人用户可通过 Gemini app 的 Spark 体验。\n\n## 我的看法\n\nGoogle 把 Flash 系列的发布节奏卷到 3 周一次,本质是在用「工作马模型高频小步快跑」反制 Anthropic Claude Opus 5、OpenAI GPT-5.6 Sol 这种「季度性大版本」路线。3.7 Flash 在真实软件工程基准上 16 个百分点的跃迁,加上价格腰斩,这两件事叠加意味着:Agent 时代的开发者心智正在从「哪个模型最强」切换到「哪个模型性价比最好、能撑得住生产部署」。Google 选择把筹码压在 Flash 而非等待 Gemini 3.5 Pro——这个选择本身,就是 Google 对模型发布策略的一次重新校准。\n\n(参考来源:[Google AI Blog 官方公告](https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Fmodels-and-research\u002Fgemini-models\u002Fintroducing-gemini-3-7-flash\u002F)、[Gemini 3.7 Flash Model Card](https:\u002F\u002Fdeepmind.google\u002Fmodels\u002Fmodel-cards\u002Fgemini-3-7-flash\u002F))","https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Fmodels-and-research\u002Fgemini-models\u002Fintroducing-gemini-3-7-flash\u002F","4d11edad-2df6-45f6-b71f-70f65de7f7fd",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"a9524a82-a7c5-4daa-bb4b-a7ee77bb0b94","gemini",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"8cf7490f-2449-4ba7-be19-61befa0d92b4","google",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"e5bf15b4-cb18-42cd-ad29-2cbd84890d12","en","Gemini 3.7 Flash: FrontierCode 43.6%, half the price","Google released Gemini 3.7 Flash on August 13, 2026, just three weeks after Gemini 3.6 Flash. The new \"most intelligent workhorse model\" jumps on coding and knowledge-work benchmarks: FrontierCode 1.1 Main 43.6% vs 34.4%, DeepSWE v1.1 65.3% vs 49.0%, WebDev Arena Elo 1588 vs 1538, GDP.pdf 34.0% vs 22.0%. Pricing is half of 3.6 Flash at $0.75 \u002F $3.75 per million input\u002Foutput tokens through end of 2026. The model exposes customizable thinking budgets and becomes the default for Gemini Spark across 160 countries.","## Three-week cadence: Google compresses Flash releases into a weekly feedback loop\n\nOn August 13, 2026, Google published Gemini 3.7 Flash on the Google AI Blog, just three weeks after Gemini 3.6 Flash shipped on July 21, 2026. Senior Director of Product Management Tulsee Doshi framed 3.7 Flash as \"our most intelligent workhorse model yet\" tuned for coding and agent workflows.\n\nA three-week gap is unprecedented in the Flash line. The July 21 batch alone dropped 3.6 Flash, 3.5 Flash Lite, and 3.5 Flash Cyber together, and 3.7 Flash immediately follows. Google cites developer feedback plus algorithmic innovation, but the cadence itself is a signal: the model's feedback loop has been compressed from quarters to weeks.\n\n## Benchmark jumps on coding and knowledge work\n\nThe headline numbers (from the official blog post [Introducing Gemini 3.7 Flash](https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Fmodels-and-research\u002Fgemini-models\u002Fintroducing-gemini-3-7-flash\u002F)):\n\n- **FrontierCode 1.1 Main**: 43.6% (3.7) vs 34.4% (3.6), +9.2 points\n- **DeepSWE v1.1**: 65.3% vs 49.0%, +16.3 points\n- **WebDev Arena Elo**: 1588 vs 1538\n- **GDP.pdf (complex document comprehension)**: 34.0% vs 22.0%, +12.0 points\n- **AutomationBench (Zapier real workflows)**: 30.4% vs 17.0%, +13.4 points\n\nDeepSWE v1.1 and AutomationBench deserve attention. DeepSWE comes from Cognition and stresses real software engineering; AutomationBench comes from Zapier and tests whether a model can replace humans in actual business workflows. The 13+ point jumps on both suggest Google is not just stabilizing Flash's coding ability — they are pushing it toward harder, more production-shaped work.\n\n## Half-price promo: another round of price compression\n\nPricing is aggressive. Google offers an introductory price of **$0.75 per million input tokens and $3.75 per million output tokens** through December 31, 2026. **That is half of 3.6 Flash**. After January 1, 2027, pricing returns to $1.50 \u002F $7.50 per million tokens.\n\nCombining a 50% price cut with benchmark gains is Google buying share in the agent era. The post lists Box, Browser Use, Cartwheel, Harvey, Hebbia, LangChain, Nunu.ai, Open Code, Pydantic, and Stanford's Department of Biology as early customers. Stanford Biology in particular signals that 3.7 Flash has credibility in knowledge-dense sciences, not just coding benchmarks.\n\n## Customizable thinking budgets\n\nThe 3.7 Flash model card calls out a specific capability: **support for customizable thinking configurations that control the trade-off between quality, cost, and latency**.\n\nThis is a deliberate departure from the \"fully automatic thinking\" pattern of Gemini 3.5 Pro. 3.7 Flash hands the thinking knob to developers. For production deployments, low-latency paths (autocomplete, UI rendering) can dial reasoning down; hard tasks (code refactors, contract review) can dial it up. The marginal cost structure becomes predictable rather than driven entirely by token spend.\n\n## Gemini Spark switches to 3.7 Flash across 160 countries\n\n3.7 Flash becomes the default for Gemini Spark starting today. Spark is Google's \"24\u002F7 personal AI agent\" launched at I\u002FO 2026, available to Google AI Pro and Ultra subscribers in 160 countries. Spark users will inherit stronger tool use and workflow orchestration without any action on their part — a tightly coupled model-and-product rollout.\n\n## Safety: CBRN and cyber offense hardening\n\n3.7 Flash ships with updated Frontier Safety safeguards covering chemical, biological, radiological, and nuclear (CBRN) misuse and cyber offense, while explicitly preserving beneficial uses. Google links to its bioresilience approach and cyber program documents. Bundling safety updates with model releases has been standard practice for frontier labs in the past 18 months, but Google front-loaded the section, reflecting ongoing regulatory and investor attention to model misuse.\n\n## Access paths\n\nDevelopers can use 3.7 Flash through the Gemini API in Google AI Studio, Android Studio, and Antigravity. Enterprises can reach it through the Gemini Enterprise Agent Platform and the Gemini Enterprise app. Individuals can experience it via Gemini Spark in the Gemini app for Google AI Pro and Ultra subscribers in supported countries.\n\n## My take\n\nCompressing Flash's release cadence to three weeks is, in effect, Google using \"small, frequent workhorse updates\" to counter the quarterly big-version cadence of Anthropic Claude Opus 5 and OpenAI GPT-5.6 Sol. Pair a 16-point jump on real software-engineering benchmarks with a price cut, and the message lands: developers in the agent era are pivoting from \"which model is the strongest\" to \"which model is the best value that can survive production load.\" Google choosing to lean into Flash rather than wait for Gemini 3.5 Pro is, in itself, a recalibration of Google's model release strategy.\n\n(Reference: [Google AI Blog announcement](https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Fmodels-and-research\u002Fgemini-models\u002Fintroducing-gemini-3-7-flash\u002F), [Gemini 3.7 Flash Model Card](https:\u002F\u002Fdeepmind.google\u002Fmodels\u002Fmodel-cards\u002Fgemini-3-7-flash\u002F))","gemini-3-7-flash-coding-agent-fast","2026-08-13T09:00:00Z","2026-08-13T20:11:10.511156Z","2026-08-19T01:48:03.231362Z",true,"agent",218,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"7b9cdf6e-5ef0-4ece-ab6c-e8cec1b02397","Google 重组 DeepMind 领导层,Gemini 研发提速应对 Anthropic 与 OpenAI 竞争","google-deepmind-reshuffle-gemini-speed","2026-08-25T07:00:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"bcedeb8e-e5eb-4bbc-98b8-ea12f869055f","Google 收编 DeepMind：25 年最大 AI 重组","google-deepmind-centralization-gemini-catchup","2026-08-14T08:00:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"4bd8e8bd-7066-4ab7-bd97-e24ea3921395","Gemini 因编程落后推迟两月:Brin 4 月督促背后,Google 把研发「收回到一个人」手里的组织账本","google-gemini-coding-behind-deepmind-reshuffle","2026-08-14T03:30:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"8b6c20ec-7222-48cf-af2c-ac97466a2b0a","Gemini 月活破 10 亿:Google 第一次把 AI 助手做成自家「最快十亿用户产品」","gemini-app-1b-monthly-users","2026-08-12T03:00:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"34edaffc-6b5c-4df1-9e2f-d864cada6063","Gemini 走进 K-12 课堂：Google 把「上下文」塞进每个作业","gemini-classroom-k12-contextualized-prompts","2026-08-07T02:00:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"df25ef84-5f0b-477b-9940-a0da477d8169","DeepMind 新主帅接棒：Hassabis 退任，Gemini 4 成 Google 筹码","deepmind-kavukcuoglu-gemini-4-reshuffle","2026-08-07T00:00:00+00:00"]