Model routing is sending each job to the AI model that fits it, by cost and difficulty. Routine, checkable work goes to a small, fast, cheap model, and the hard calls go to the strongest model you can get. Most AI products already route behind the scenes. For a small business it becomes your decision once you stop paying a flat monthly fee and start paying for each request.
An AI model is the trained engine that reads your words and writes the reply. Every major lab sells several sizes of them, and you'll also hear the term LLM routing, which means the same thing (an LLM, or large language model, is the kind of model behind ChatGPT and Claude). The app you type into is a separate layer, and a separate post covers the program wrapped around the model and how it shapes what you see.
Who already does this
An accounting firm routes work every day without calling it anything. A staff accountant prepares the routine return, and a partner reviews the one with real exposure. You wouldn't bill partner hours to put a date on a form, and you wouldn't hand the newest hire the unusual audit and walk away.
The big AI apps make the same call for you. When OpenAI launched its GPT-5 family in 2025, ChatGPT began sending each question either to a fast model or to a slower one that works through a problem step by step. Then it pulled back. The Decoder reported OpenAI's router rollback in December 2025. Free users went back to the fast model by default, partly because the slow one made people wait and partly because it cost more to run, while Plus and Pro subscribers kept automatic routing.
Microsoft took a different path. Microsoft 365 Copilot now offers Anthropic's Claude alongside OpenAI's models, and an administrator controls whether Anthropic's models are available to the organization.
So a question like "which model does Copilot use?" has a moving answer. It depends on the request, the plan, and the settings. That's fine while you pay a subscription, because the bill doesn't change with the model.
Routing by cost and difficulty
The bill changes the day you pay per use. That happens when a business connects a model to its own systems through an API, the doorway one program opens so another program can send it work (the post on APIs, connectors, and MCP explains the doorways). From then on you pay for tokens, the small chunks of text a model reads and writes, priced per million.
Here's how far apart the tiers inside one family sit. Anthropic's published API pricing and Google's Gemini API pricing, both read on Oct. 17, 2026, list these output prices per million tokens for their current models.
The numbers
| Tier | Output price |
|---|---|
| Smallest | $0.50 |
| Middle | $10 |
| Large | $20 |
| Largest | $50 |
Output price per million tokens. The smallest tier's price applies to prompts up to 100,000 tokens.
Source: Anthropic API pricing, claude.com/pricing, read Oct. 17, 2026
The numbers
| Tier | Output price |
|---|---|
| Smallest | $2.50 |
| Middle | $3.75 |
| Largest | $10 |
Output price per million tokens, paid tier, current models that aren't in preview. Google's page lists the middle tier at $3.75 through Dec. 31, 2026 and $7.50 starting Jan. 1, 2027, and the largest at $10 through Nov. 30, 2026 and $20 starting Dec. 1, 2026.
Source: Gemini API pricing, ai.google.dev/gemini-api/docs/pricing, read Oct. 17, 2026
That's a hundredfold spread between the bottom and the top of Anthropic's lineup, and Google's runs fourfold. Put rough numbers on it. An outdoor outfitter writing 3,000 product descriptions a month, at around 400 tokens each, produces about 1.2 million tokens of output. Through the largest model that's roughly $60 a month in output alone. Through the smallest it's about 60 cents. Input tokens cost extra at every tier, and the shop would still need to read what comes back, but the ratio holds.
Difficulty pulls the other way. Smaller models make more mistakes on work with conditions inside conditions, like reconciling a messy month of transactions or drafting a contract clause where one exception changes three others. You pay twice for an answer you have to redo. The work that suits a small model is short, frequent, and easy to check, and a wrong answer there costs you a glance.
What goes where
In rough terms, tagging emails and reformatting a price sheet go to the smallest model, routine drafts a customer will read go to a small or middle one, and anything with legal, tax, or safety consequences goes to the largest, with a person signing off.
A router can be as plain as a rule you write ("summaries go here, quotes go there") or a small model whose only job is to read the request and pick the destination. Some routers escalate, trying the cheap model first and passing the job up when the answer looks shaky. That last design is clever, and it's also where I'd be most careful, because "looks shaky" has to be defined by someone, and a confident wrong answer doesn't look shaky at all.
When routing is worth your time
If you use AI through a subscription to ChatGPT, Claude, Gemini, or Copilot, you mostly don't need to think about this. Your vendor does the routing and the price stays flat. The question gets real at volume, which usually means an automation that runs without you or an AI agent taking steps with tools, where hundreds of small requests a day add up.
The habit worth building is to look at what model a tool defaults to before it runs a thousand times. Developers and no-code tools often ship with the newest, largest model selected, because that makes the demo look good.
Which model family you route between is a separate choice, and frontier versus open-source models covers rented top capability against weights you can run yourself. The pillar on how to use AI in your business argues for the lowest level of AI that does the job, and routing is that same argument at the level of a single request. If you'd like help sorting your own jobs into the table above, that's part of our intelligence work.