--- title: "Model Routers: Who Picks the Model" description: "Enterprises should no longer run on a single model, and the question is who decides which model answers a given prompt. A frontier model and a small open-weight model can differ in price by orders of magnitude, yet most production workloads still send every request to the same default. The market has responded with a wave of products sold under labels used interchangeably: gateway, router, proxy, load balancer. They are not the same thing." image: https://alten.capital/hubfs/Covers-2.png --- [Skip to content](https://alten.capital/blog/model-routers-who-picks-the-model#main-content) [![Alten-Capital-Logo-white-2](https://alten.capital/hubfs/Alten-Capital-Logo-white-2.svg "Alten-Capital-Logo-white-2")](https://alten.capital) - [About us](https://alten.capital/about-us) - [Strategy](https://alten.capital/strategy) - [Insights](https://alten.capital/blog) - [Contact us](https://alten.capital/contact-us) Search Search - There are no suggestions because the search field is empty. [Alten Capital](https://alten.capital/blog/author/alten) October 15, 2026 # Model Routers: Who Picks the Model Enterprises should no longer run on a single model, and the question is who decides which model answers a given prompt. A frontier model and a small open-weight model can differ in price by orders of magnitude, yet most production workloads still send every request to the same default. The market has responded with a wave of products sold under labels used interchangeably: gateway, router, proxy, load balancer. They are not the same thing. --- The space is best understood as a ladder, with each level handing more of the model decision from the developer to software: | **Category** | **What it does** | **Examples** | | --- | --- | --- | | Integration gateways | One API across providers with key management, caching, failover and spend control; model choice follows rules the developer writes | LiteLLM, Portkey, Helicone, Cloudflare AI Gateway, Vercel AI Gateway | | Aggregators | A marketplace with one API and one bill across hundreds of models, increasingly with an automatic model selector on top | OpenRouter | | Platform-native routers | Per-prompt model selection, but only within a single cloud's own catalog | Azure AI Foundry Model Router, Amazon Bedrock Intelligent Prompt Routing | | Learned routers | A trained model predicts which LLM will answer each prompt best at the lowest cost | Not Diamond, Martian (commercial); RouteLLM, vLLM Semantic Router (open source) | The first two levels are plumbing. They carry most production traffic, but when they route, they route by instruction: send this workload to model X, fall back to Y if it fails. Platform-native routers make a genuine per-prompt decision, but inside a walled garden; Bedrock's router, for example, currently chooses only between tiers of the same model family. vLLM Semantic Router uses a small classifier to read the intent of each prompt, picking the cheapest capable model and deciding whether reasoning is needed at all. RouteLLM learns from human preference data when a cheaper model's answer is good enough. Not Diamond and Martian predict output quality for each candidate model before the call is made. A concrete case can show the economics. Picture an AI agent resolving a customer support ticket. Classifying the ticket and retrieving the account history can go to a small, fast model; drafting a code fix goes to a frontier model; the final customer reply goes back to a mid-tier model. Without a router, the whole workflow would run on the most expensive model by default. The market's view so far is that value accrues to whoever owns the traffic, not to whoever has the smartest algorithm. OpenRouter was processing 25 trillion tokens a week when it raised at roughly $1.3 billion in May, and three months later Stripe agreed to acquire it for reportedly more than $7 billion. Tellingly, OpenRouter's automatic model selection is powered by Not Diamond rather than built in-house. The router is becoming a feature of the platform that sits on the traffic, which is the same pattern that shaped the shift toward open-weight models flowing through these platforms (we wrote about model openness here [https://alten.capital/blog/models-closed-open-weight-open-source](https://alten.capital/blog/models-closed-open-weight-open-source)) For technology services firms, the implication is that routing is a delivery capability, not a product purchase. A router is only as good as the evals and context engineering behind it. A generic router has no way of knowing that a client's claims-processing prompts need a specific model, or that a cheaper model fails quietly on its contract language. Building those routing policies, the evaluation sets that validate them, and the context that informs them is client-specific work, and it sits naturally inside the orchestration layer of a human-agent delivery model (we wrote about that layer here: [https://alten.capital/blog/the-ai-agent-orchestrator](https://alten.capital/blog/the-ai-agent-orchestrator)). Firms that bill on outcomes keep the savings that good routing creates; firms that bill on time hand them to the client. Alten Capital invests in exceptional management teams to accelerate high-growth technology services businesses. Reach out to explore partnership opportunities. ![Alten-Capital-Logo-white-3](https://alten.capital/hubfs/Alten-Capital-Logo-white-3.svg "Alten-Capital-Logo-white-3") We invest in exceptional management teams to accelerate high-growth technology services businesses. --- [Let's conect](https://www.linkedin.com/company/alten-capital/) --- [Home](https://alten.capital) [About Us](https://alten.capital/about-us) [Strategy](https://alten.capital/strategy) [Insights](https://alten.capital/blog) [Contact us](https://alten.capital/contact-us) --- New York Office 386 Park Ave South ,18th Floor New York, NY 10016 --- Subscribe to the latest news in technology services investing and M&A. © 2023 - Alten Capital - All rights reserved.   ```json { "@context" : "https://schema.org", "@type" : "BlogPosting", "author" : { "@type" : "Person", "name" : "Alten Capital", "url" : "https://alten.capital/blog/author/alten" }, "dateModified" : "2026-10-15T14:18:55.024Z", "datePublished" : "2026-10-15T14:18:55.000Z", "headline" : "Model Routers: Who Picks the Model", "image" : [ "https://alten.capital/hubfs/Covers-2.png" ], "mainEntityOfPage" : { "@id" : "https://alten.capital/blog/model-routers-who-picks-the-model", "@type" : "WebPage" }, "publisher" : { "@type" : "Organization", "logo" : { "@type" : "ImageObject", "url" : "https://alten.capital/hubfs/Alten-Capital-Logo.svg" }, "name" : "Alten Capital" } } ```