Cursor 发布智能模型路由系统 Cursor Router

今天我们推出了 Cursor Router，这是我们为团队和企业打造的智能模型路由器。

Cursor Router 让团队可以自动将每个请求路由到最适合完成该任务的模型，以较低的成本提供前沿智能。

我们在数千名企业开发者的生产流量中观察到了非常强的成果。在与数十家企业的早期访问期间，客户以大约低 30–50% 的成本获得了前沿性能。

在数百万次请求的在线 A/B 测试中，Cursor Router 实现了前沿质量表现，同时节省了 60% 的成本。

Cursor 每周处理数亿次编码请求，覆盖每个模型和提供商，并对用户偏好和代码库中的内容具有独特的可见性。模型中立性一直是 Cursor 工作的核心，今天我们将这些数据和专业知识应用于你的团队。

通过 Cursor Router，我们的目标是为团队提供每项任务的最佳性能和体验，而无需花费超出工作所需的成本。

大约 60% 使用 Cursor 的开发者选择单个模型作为日常使用工具。这导致日常工作以前沿价格完成，而 AI 支出比产出质量增长得更快。Cursor Router 通过在模型运行前对每个请求进行分类来解决这一问题。

从本质上讲，Cursor Router 是一个分类器，根据用户查询将其路由到最佳模型选项。我们在超过 60 万个实时请求上训练了 Cursor Router，并在数百万个由 Cursor Router 指导的实时请求上进行了在线 A/B 测试，以用户满意度（AFC）为奖励优化性能。

Cursor Router 根据查询、上下文、任务复杂性和领域分析每个请求，并结合我们对每个模型行为的了解。我们学习每个模型最擅长的内容，并将其路由到最有效的选项。简单工作由最具性价比的模型处理，UI 更新由最具品味的模型处理，更复杂、长远的问题由前沿推理模型处理。

我们设计了路由分类器，以适应不断早期和频繁发布更新模型的世界。这样，当更新、更强大的模型发布时，我们可以轻松更新 Cursor Router，让体验持续改进。

Cursor Router 在训练和评估中都具有缓存感知能力。它在一个会产生缓存未命中的数据集上进行训练，并在生产环境中进行评估，我们报告的成本节省包括路由决策中缓存未命中的成本。

Cursor Router 有三种模式：Intelligence、Balance 和 Cost，可以让你调整自己在成本-智能帕累托前沿的位置。

我们发现，自动 Intelligence 模式在用户对输出满意度方面接近 Fable，同时在团队成本上降低约 60%，并且在几乎相同的成本下将满意度提升约 15% 相比 Opus 4.8。

类似地，自动 Balance 模式在用户对结果的满意度上高于 Opus 4.8，同时成本降低约 36%。与 GPT-5.6 Sol 相比，自动 Balance 在较低的花费下提供相当的满意度。

我们选择使用大型在线 A/B 测试来衡量路由器的有效性，而不是离线评估。虽然离线评估是质量的有用代理，但它们受限于数据量小、与实际使用的距离以及将成功减少到评分标准的难度。

离线评估还忽略了由模型切换带来的额外缓存未命中成本。真实的路由发生在一段对话中：选择哪个模型、何时切换。

在线 A/B 测试将 Cursor Router 放在真实世界中进行测试，涵盖数百万个任务和对话。工程师编写代码、提出后续问题、遇到错误并继续工作，通常一周内会处理上百个请求。这些都是模型路由器需要良好表现的环境。

在输出质量方面，我们测量了：

在过去九个月中，我们依靠这些指标评估每次模型发布和功能改进。

在过去两周，Cursor Router 对一部分企业客户进行了早期访问。我们将他们实际支付的费用与相同流量按 Opus 4.8 API 收费的费用进行比较。

在早期访问中，三个高流量账户的几千名用户在自动路由请求上节省了 30%–50%，而无需降低质量，这相比将所有请求路由至 Opus 4.8。

每个请求的成本只是故事的一半。工程领导关心这些节省是否体现在实际交付工作中，因此我们观察了每次提交的成本，这一模式依然成立。

对于单次提交，我们观察到在智能模式下，Cursor Router 每次提交的成本为 6.76 美元，在平衡模式下为 4.63 美元。

GPT-5.6 Sol 的成本与智能模式相当，但输出的用户满意度较低。与此同时，Fable 5 和 Opus 4.8 的提交成本较 Cursor Router 有所溢价，分别为 12.69 美元和 7.34 美元。

这种差距就是路由的实际案例。Cursor Router 将困难任务保留给最有能力的模型，而将常规工作移出前沿定价。

Cursor 针对团队和大型组织设计了我们的路由器。路由器使用数据驱动的分类法，同时管理员和终端用户仍然可以选择其在成本-智能帕累托前沿的位置。

在模型选择器中选择自动模式，并从三种优化模式中选择，使你沿前沿移动：

管理员可以决定 Cursor Router 如何在团队中部署。你可以按团队或小组启用它，选择成员可以选择的模式，设置默认值，并允许或阻止特定模型。

Cursor Router 是 Cursor 推动代币效率的一个部分。选择正确的模型只有在代理本身保持精简时才有意义，因此我们继续削减肉工具中的浪费：https://cursor.com/blog/continually-improving-agent-harness。

动态工具调用：https://cursor.com/blog/dynamic-context-discovery 是另一个明显的例子，其中大多数本地工具描述不再加载到每个提示中。模型在首次需要时查找它们，这遵循我们已经在 MCP 中使用的相同模式。这使常用工具（如读取和编辑）保持热状态，而不常用工具只有在代理实际调用时才进入提示。

除了 Cursor Router，我们还持续提高模型池的下限和上限：Grok 4.5：https://cursor.com/blog/grok-4-5 扩大了 Cursor Router 在更困难、高成本工作中可调用的范围。Composer 在日常路径上不断改进，因此较低成本的结果保持接近前沿质量，而无需支付前沿价格。

Cursor Router 现已在桌面、网页、iOS、CLI 和我们的 SDK 上对团队和企业计划开放。

阅读我们的文档：https://cursor.com/docs/cursor-router 以及更新日志：https://cursor.com/changelog/router。

归档于：产品：/博客/主题/产品

使用 iOS 版 Cursor 随时随地构建

为 Cursor Enterprise 推出组织功能

Today we're launching Cursor Router, our intelligent model router for teams and enterprises.

Cursor Router lets teams automatically route every request to the most capable model for the task, delivering frontier intelligence at a lower cost.

We've observed incredibly strong results on production traffic across thousands of enterprise developers. During our early access period with dozens of enterprises, customers got frontier performance at approximately 30–50% lower cost.

In online A/B tests across millions of requests, Cursor Router delivered frontier-quality performance at 60% savings.

Cursor routes hundreds of millions of coding requests each week across every model and provider, with unique visibility into what users like and what stays in the codebase. Model neutrality has always been core to how Cursor works, and today we're putting that data and expertise to work for your team.

With Cursor Router, our goal is to provide teams with the best performance and experience for every task, without spending more than the work requires.

Roughly 60% of developers using Cursor pick a single model as their daily driver. This results in routine work being completed at frontier prices, and AI spend growing much faster than output quality. Cursor Router fixes that by classifying each request before a model runs.

At its core, Cursor Router is a classifier that routes users to the best model option based on their query. We trained Cursor Router on 600k+ live requests and evaluated performance in an online A/B test across millions of live requests directed by Cursor Router, optimizing for user satisfaction (AFC) as a reward.

Cursor Router analyzes each request on query, context, task complexity, and domain, combined with what we know about each model's behavior. We learn what each model is best at, and route to the most effective option. Simple work goes to the most price-efficient models, UI updates go to the model with the best taste, and more complex, long-horizon problems go to frontier reasoning models.

We designed our routing classifier for a world in which updated models get shipped early and often. This way as newer and more powerful models are released, we can easily update Cursor Router, so the experience keeps improving.

Cursor Router is cache-aware in both how it is trained and evaluated. It is trained on a dataset where routing results in cache misses, and evaluated in production where our reported cost savings include the cost of cache misses in routing decisions.

Cursor Router has three modes: Intelligence, Balance, and Cost which let you adjust where you are on the cost-intelligence Pareto frontier.

We found that Auto Intelligence mode lands near Fable on user satisfaction of output at about 60% lower cost for teams , while also lifting satisfaction about 15% over Opus 4.8 at nearly the same cost.

Similarly, Auto Balance lands above Opus 4.8 on user satisfaction with the results at about 36% lower cost. Against GPT-5.6 Sol, Auto Balance delivers comparable satisfaction at a lower spend rate.

We chose to measure the efficacy of our router using large online A/B tests instead of offline evals. While offline evals are useful proxies for quality, they're limited by their small size, their distance from real-world usage, and the difficulty of reducing success to a rubric.

Offline evals also omit the extra cache-miss cost that comes from switching models. Real routing happens across a conversation: which model to pick, and when to switch.

Online A/B tests put Cursor Router to test in the real world across millions of tasks and conversations. Engineers write code, ask follow-ups, hit errors, and keep going, often across hundreds of requests in a week. Those are the conditions under which a model router needs to perform well.

In terms of quality of output, we measured:

We have relied on these metrics to evaluate every model launch and harness improvement in the past nine months.

Over the past two weeks, Cursor Router has been in early access with a selection of enterprise customers. We compared what they actually paid against the same traffic, priced entirely at Opus 4.8 API rates.

In early access, three high-volume accounts with thousands of users saved 30%–50% on Auto-routed requests versus routing everything to Opus 4.8, with no decrease in quality.

Cost per request is only half the story. Engineering leaders care whether those savings show up in real shipped work, so we looked at cost per commit, and the pattern held.

For a single commit, we observed Cursor Router had a lower cost per commit of $6.76 for Intelligence mode and $4.63 for Balance.

GPT-5.6 Sol matched the cost of Intelligence but had lower user satisfaction with the output. Meanwhile, Fable 5 and Opus 4.8 produced commits at a cost premium to Cursor Router at $12.69 and $7.34 respectively.

That gap is the practical case for routing. Cursor Router keeps hard tasks on the most capable models and moves routine work off of frontier pricing.

Cursor designed our router with teams and large organizations in mind. The router uses a data-driven taxonomy, while admins and end users can still choose where it sits on the cost-intelligence Pareto frontier.

Select Auto mode in the model picker, and choose from three optimization modes that move you along the frontier:

Admins can decide how Cursor Router rolls out across teams. You can enable it per team or group, choose which modes members can select, set the default, and allow or block specific models.

Cursor Router is one piece of how Cursor drives token efficiency. Choosing the right model only matters if the agent itself stays lean, so we keep cutting waste in the harness：https://cursor.com/blog/continually-improving-agent-harness around it.

Dynamic tool calling：https://cursor.com/blog/dynamic-context-discovery is another clear example where most native tool descriptions are no longer loaded into every prompt. The model looks them up the first time it needs them, following the same pattern we already use for MCPs. This keeps common tools like read and edit hot while less commonly used tools only enter the prompt when the agent actually calls them.

Alongside Cursor Router, we keep raising the floor and the ceiling of the model pool: Grok 4.5：https://cursor.com/blog/grok-4-5 widens what Cursor Router can draw from on harder, higher-cost work. Composer keeps getting better on the everyday path, so lower cost turns stay close to frontier quality without paying frontier prices.

Cursor Router is available today for Teams and Enterprise plans across desktop, web, iOS, CLI, and our SDK.

Read more in our docs：https://cursor.com/docs/cursor-router and changelog：https://cursor.com/changelog/router.

Filed under: product：/blog/topic/product

Build from anywhere with Cursor for iOS

Introducing organizations for Cursor Enterprise