模型更新太快，你的工程团队跟不上了。
8月19日，以企业支出管理闻名的Ramp做了一件出人意料的事：把内部运行了3年的AI模型路由系统对外开放，取名Router.com。一个API接入，智能调度OpenAI、Anthropic、SpaceXAI（Grok）等多家模型，自动为每个请求选择性价比最优的模型。Ramp称现有客户平均节省40%的推理成本，且不牺牲质量和稳定性。2026年底前路由功能免费，还送26美元模型信用额。
就在上周，Stripe刚以70亿美元收购OpenRouter（039期已报道）。支付公司和费用管理公司，不约而同杀入了AI"交通指挥"赛道。
模型大爆炸时代，企业需要一个AI调度员
问题很简单：2026年的大模型市场已经乱成一锅粥。OpenAI发了GPT-5.6系列，Anthropic出了Claude Opus 5和Sonnet 5，Google有Gemini 3.7，xAI有Grok 4.6，开源阵营有Llama 4、Qwen3.8、DeepSeek V4……每家都说自己最强，每个都在不同任务上各有胜负，每个月都有新模型发布、新降价、新benchmark。
对企业工程团队来说，这是噩梦。你花两周时间评测完选定了一个模型，下周另一家降价30%；你刚把工作流迁到Claude，发现某个特定任务Luna更快更便宜；你写死了调用GPT-5.6 Sol的代码，结果Sol在高峰期限速，服务直接挂掉。Ramp的AI Index数据显示，企业AI支出自2025年6月以来增长了20.7倍，但大部分公司根本不知道自己的钱花在了哪里、哪些请求其实可以用更便宜的模型处理。
Router的解法是四大策略：Flex tier自动在弹性层（同模型但延迟不稳定、价格更低）和标准层之间切换，弹性层延迟达标时自动走便宜路径；Shadow models把真实流量的副本发给候选模型做对比测试，不影响生产就能评估新模型；Benchmark routing让开发者定义benchmark权重，Router自动选得分最高的模型；NVIDIA Switchyard在Agent多步任务中动态切换，简单步骤用小模型、难步骤才升级到旗舰模型。
Ramp举了一个内部例子：一个编码Agent任务，58轮对话用Luna处理，只有最后5轮难的升级到Sol，两个模型的prompt缓存都保持预热状态。既拿到了旗舰模型的能力，又没有为每一轮都付旗舰模型的价格。
支付公司为什么在做AI基础设施？
Ramp是做企业费用报销和支出管理起家的，Stripe是做支付网关的。这两家公司为什么同时盯上了模型路由？
答案藏在它们的业务基因里。Ramp和Stripe的核心能力都是在复杂系统中做智能路由。Stripe在全球支付网络里决定每笔交易走哪个支付通道成功率最高、手续费最低；Ramp在企业采购流程里决定每张账单走哪个审批路径最合规、最省钱。模型路由本质上是同一个问题：每一个AI请求应该走哪个"通道"（模型），在质量、速度、成本之间找最优解。
Ramp强调自己的独立性："我们不做模型，所以没有理由偏向任何一家。"这跟Stripe收购OpenRouter的逻辑一致——做中立的AI流量调度层，谁的模型好、价格优，就把流量给谁。Ramp还提供详细的请求级成本可视化：每个请求用了什么模型、花了多少钱、延迟多少、有没有fallback——让企业第一次能看清自己的AI账单。
"AI是大多数公司增长最快的支出项，也是最算不清账的一项。Router让你第一次知道每一分AI钱花在了哪里。"—— Veeral Patel, Ramp Applied AI负责人
模型路由这个赛道正在快速成型。微软Foundry做模型市场，Stripe/OpenRouter和Ramp/Router做流量调度，Perplexity也在8月21日发布了Agent API（单一端点接入41个模型）。当模型本身成为可替换的商品，路由层就是新的流量入口和价值捕获点。
对企业来说这是好事：竞争意味着价格下降、工具成熟。对模型公司来说这是警示：如果路由层可以随时把流量从你这里切到竞争对手那里，你的护城河到底是什么？
明天见。
Models ship faster than your engineering team can evaluate them.
On August 19, Ramp — best known for corporate spend management — did something unexpected: it open-sourced the AI model routing system it had run internally for three years under the name Router.com. One API integration intelligently orchestrates models from OpenAI, Anthropic, SpaceXAI (Grok), and others, automatically selecting the best cost-performance option for each request. Ramp says existing customers save an average of 40% on inference costs without sacrificing quality or uptime. Routing is free through the end of 2026, plus $26 in model credits.
Just last week, Stripe acquired OpenRouter for $7 billion (covered in Issue 039). A payments company and a spend management company — both piling into AI "traffic control."
The Model Cambrian Explosion Creates a Need for an AI Dispatcher
The problem is simple: the 2026 LLM market is chaos. OpenAI shipped the GPT-5.6 family; Anthropic released Claude Opus 5 and Sonnet 5; Google has Gemini 3.7; xAI has Grok 4.6; the open-source camp has Llama 4, Qwen3.8, DeepSeek V4… Everyone claims to be the best, each wins on different tasks, and every month brings new models, new price cuts, new benchmarks.
For engineering teams, this is a nightmare. You spend two weeks evaluating and pick a model; next week a competitor cuts prices 30%. You migrate a workflow to Claude and discover a specific task runs faster and cheaper on Luna. You hard-code GPT-5.6 Sol calls and the service goes down during peak rate limits. Ramp's AI Index shows enterprise AI spending has grown 20.7x since June 2025, yet most companies have no idea where their money is going or which requests could be handled by cheaper models.
Router's answer is four strategies: Flex tier automatically switches between discount-priced flex capacity (same model, less predictable latency) and standard tiers, routing to the cheaper path when flex latency meets SLAs; Shadow models duplicate real traffic to candidate models for head-to-head comparison without touching production; Benchmark routing lets developers define benchmark weightings so Router auto-selects the top-scoring model; NVIDIA Switchyard dynamically escalates within multi-step agent tasks — keeping routine turns on smaller models and only escalating to frontier models for hard steps.
Ramp shared an internal example: a coding agent task where 58 turns were handled by Luna and only the final five hard turns escalated to Sol, both models' prompt caches staying warm. Frontier capability where it matters, without paying frontier prices for every single turn.
Why Are Payments Companies Building AI Infrastructure?
Ramp built its business on expense reporting and corporate spend management. Stripe built its on payment gateways. Why are both going after model routing?
The answer lies in their business DNA. Both Ramp and Stripe's core competency is intelligent routing in complex systems. Stripe decides which payment rail gives the highest approval rate and lowest fees for every transaction across global networks. Ramp decides which approval path is most compliant and cost-effective for every invoice. Model routing is the same problem: which "rail" (model) should handle each AI request, optimizing for quality, speed, and cost?
Ramp emphasizes its independence: "We don't build models, so we have no reason to favor any particular one." That mirrors Stripe's logic with OpenRouter — be the neutral AI traffic control layer, sending volume to whoever has the best model at the best price. Ramp also provides per-request cost observability: model used, price, latency, fallback attempts — giving enterprises their first clear view of the AI bill.
"AI is the fastest-growing line item at most companies, and the one they can least measure. Router gives you the first real look at where every AI dollar goes."— Veeral Patel, Head of Applied AI at Ramp
The model routing category is taking shape fast. Microsoft Foundry runs the model marketplace; Stripe/OpenRouter and Ramp/Router handle traffic orchestration; Perplexity also launched an Agent API on August 21 (single endpoint accessing 41 models). When models themselves become interchangeable commodities, the routing layer becomes the new gateway and value capture point.
That's good for enterprises: competition means lower prices and more mature tooling. For model companies it's a warning: if a routing layer can shift traffic away from you to a competitor at any moment, what exactly is your moat?
See you tomorrow.
当模型成为可替换商品，路由层就是新的流量入口和价值捕获点。
—— Dawn Vision编辑部
When models become interchangeable commodities, the routing layer becomes the new gateway and value capture point.
— The Dawn Vision Editorial Desk
Ramp · Router.com · 模型路由 · 40%成本节省 · 四大策略 · NVIDIA Switchyard · Stripe OpenRouter · AI流量控制 · 企业AI支出
Ramp · Router.com · model routing · 40% cost savings · four strategies · NVIDIA Switchyard · Stripe OpenRouter · AI traffic control · enterprise AI spending
