AI model routing, send simple tasks to cheaper models

AI model routing applies the mail room reflex to models. Each letter goes to the right slot instead of everything being sent by registered post. Simple requests go to the budget model and demanding ones go to the higher tier. The bill follows suit. The quality of the work doesn't suffer anywhere along the way.

The topic has a technical reputation, with gateways, classifiers and cascades. In reality it starts much lower down. Useful routing can be set up by hand, with one decision and zero lines of code. What follows goes from principle to practice, with the three ways to route, the task to model table and the traps that wait for setups that are too ambitious.

01Key takeaways

• Routing means sending each task to the model tier that is enough, instead of paying flagship prices for everything.
• Routine requests usually make up most of everyday volume and a budget model does the job on them.
• There are three methods in order of complexity, which are fixed rules, a cascade with fallback and an automatic classifier.
• Manual routing by task type already captures the vast majority of the gain. Tools come later, if the volume justifies them.
• The quality check is to test each move down a tier blind on your own tasks before making it permanent.

02Paying flagship rates for trivial tasks

Public price lists span two orders of magnitude between the floor and the top. Rewording an email at the price of a frontier model is like sending a postcard as a tracked parcel. Real usage leans heavily toward simple work, such as sorting, rewording, short summaries and routine replies. On all those tasks the budget tier turns in equivalent work for a fraction of the price.

That logic of fitting the model to the job runs through the whole AI without subscription approach. Paying for real usage only makes sense if the usage itself is right-sized. Routing is the central tool for that, placed just behind the choice of model by the eight cost levers. The exact gaps can be read in the providers' compared price lists.

03The principle, three ways to route

Rule-based routing, the list on the wall

The simplest method fits on one sheet of paper. A given task type goes to a given model, decided once and applied every time. Drafts and rewording go to the budget tier, while sensitive summaries and final texts go to the higher tier. It needs no tool, adds no latency and only calls for one review per quarter. It's the routing that suits the vast majority of freelancers and small teams, today as much as two years from now.

The cascade, start small and escalate if needed

A cascade first tries the budget model and only moves up to the premium one if the result disappoints, either on review or through an automatic check. The saving comes from the fact that most requests stop at the first tier. The price to pay is a second call on the escalated cases, so a little delay and a double cost on that fraction. The cascade shines on volumes processed offline, such as content generation or bulk processing, where the delay of an escalation bothers nobody. It suits live dialogue less well, because every second of waiting is felt.

The classifier, a machine that sorts for you

At the top of the sophistication scale, a small model examines each request and sends it on according to its estimated difficulty. It's the heart of professional gateways and self-learning routers. It's also a project that only pays off at high volume, because the classifier itself costs money, needs tuning and has to be monitored. Remember that it exists, but don't make it your first step. The sign that it's becoming relevant is a rule sheet that no longer works because difficulty varies within a single task type, at thousands of requests per day.

04Which model for which task, the starter table

The table gives a reasonable starting split, to adjust based on your results. Tier names stay generic on purpose, because each provider's floor and flagship models change with every release. The ranking of models by price keeps the mapping up to date, so you can check which model currently fills each tier.

Task type	Suggested tier	Why
Sorting, extraction, formatting	budget	mechanical task, invisible gap
Rewording, emails, short summaries	budget	routine where the floor excels
Content drafts	budget or mid-tier	a human goes over it afterward
Texts published as is	mid-tier or premium	the cost of rework exceeds the gap
Long analysis, reasoning, sensitive code	premium	quality is worth its price here

Your personal boundary is found through a blind test, with the same task run on two tiers and the results judged without labels. Choosing the best value model gives the full protocol. Ten minutes per task type are enough to calibrate the routing sheet on your own standards rather than on reputations.

05Doing it in practice, from simplest to most tooled

By hand, switching models depending on the task

Manual routing only requires a tool that lets you pick the model at the moment of use. The rule sheet does the rest, with one second before each job to select the planned tier. The habit takes a week to form. The provider's statement shows the effect from the following month on, line by line and model by model. Costing a single call even lets you predict the gain before you start.

Through a gateway, one entry point for every model

Multi-provider gateways expose hundreds of models behind a single key. They allow split rules and even an automatic fallback when a provider is overloaded. The single OpenRouter key for many models is the best-known route. Count its pass-through fees in the math, since they're the price of all-in-one flexibility. The benefit grows with the number of providers you use. With a single model the gateway adds almost nothing, while with five models spread across three companies it replaces five consoles with one.

In a BYOK workspace, routing without thinking about it

A workspace built for personal keys puts routing into the interface. Each tool lets you pick the model from a menu. Your keys stay yours. OpenTahr works along these lines and the account is free. You choose the model per tool, so drafts can go to the budget tier and final versions to the tier you trust. The tool's plan stays fixed while usage is billed at the provider's list price. It's rule-based routing made comfortable. The rule lives in the tool's settings instead of on a sheet of paper, so it applies even on tired days.

06A month of routing, from start to finish

Take a writer who produces a hundred jobs a month, with sixty rewordings and outlines, thirty article drafts and ten summaries published as is. With everything on premium, the bill carries a hundred times the top rate. Routed according to the starter table, ninety jobs move down to the budget tier, billed at a fraction of the premium price on every public price list. Only the ten summaries keep the full rate.

The result can be read without a calculator. Most of the volume shifts to the cheapest tier and the bill follows in the same proportion, while the ten sensitive jobs keep exactly the same quality as before. The detailed figures can be rebuilt in ten minutes with your own rates. The mechanics don't change from one profile to another. Only the proportions move.

The same reasoning holds for a customer assistant or a chain of automations. Count the routine share of your volume, because it sets the ceiling on your gain. You can measure yours by reading back through a week of history, with a pencil in hand and three columns on a sheet.

07The three beginner mistakes in routing

The first mistake is routing by affection. Keeping your favorite model everywhere because you know it well is the opposite of routing, since the choice should follow the task rather than the habit. The second mistake is routing in the middle of a conversation. Switching tiers mid-dialogue loses the benefit of the input cache and muddles the consistency of the replies, so route at the start of a job, not halfway through.

The third mistake is routing without a safety net. Moving a whole category of tasks to the floor model on the same day, with no blind test and no observation period, turns the smallest weakness of the small model into a string of manual fixes. A careful switch happens task by task, with a week of observation for each. The statement and the observed quality validate each step before the next one.

08Routing as a team, when the sheet becomes a policy

With several people, routing stops being a personal habit and becomes a shared rule. That means a common sheet of tiers by task type, decided together and reviewed on the same schedule. Without it, everyone routes according to their own feel and the statement becomes unreadable, because the gaps between colleagues hide the gaps between settings.

In practice this runs through the keys, with one key per person or per client. Each key carries the routing policy in the settings of the shared tool. The provider's statement then splits naturally. Quarterly reviews rest on clean figures. The discussion is about the rules, never again about impressions or personal taste in models.

09The traps of latency, quality and drift

Three points call for care. Latency comes up first, since a cascade adds a second call on escalated cases and a classifier adds its own check to every request. For interactive use, rule-based routing stays the most responsive. Quality matters too, because moving down a tier without a blind test ends in manual rework, the hidden cost that eats the displayed gain.

Drift is the last point. A routing sheet written once and then forgotten ages badly, as models change several times a year. A quarterly review is enough, ideally backed by the usage statement. If the premium share creeps up with no business reason, the rule has slipped somewhere and deserves to be rewritten. Living routing gets reviewed. Frozen routing decays.

10Setup checklist

1. List your five most frequent AI tasks and their rough share of your volume.
2. Assign each task to a tier with the starter table, then blind test the planned moves down a tier.
3. Write the rule sheet with one line per task and keep it visible where you work.
4. Add caching for repeated input and leaner input templates, since the three levers stack.
5. Read the statement the following month, note the gain and review the sheet every quarter.

Nothing on this list requires a paid tool or engineering skills. Routing is a sorting discipline first and only then a question of architecture. Start with the sheet on the wall and move up when the volume calls for it, because the gain comes from the habit long before it comes from the tooling.

11Frequently asked questions

What is AI model routing?

It's the practice of sending each request to the model suited to its difficulty. Routine tasks go to a budget model and demanding tasks go to a premium one. Sorting is done with fixed rules, with a cascade and fallback or with an automatic classifier, depending on the volume involved.

How much does routing save?

The saving grows with the share of routine requests in your volume. Your own figure depends on your real split. You can predict it by costing two or three typical tasks on two tiers, then confirm it on the statement of your first routed month.

Does routing hurt quality?

It doesn't if each move down a tier passes the blind test before you adopt it, with the same instruction on two models and results judged without labels. Tasks that fail the test stay on the higher tier. Those that pass get the same work for a fraction of the price. Nobody sees the difference.

Which model should you choose for which task?

Use the budget tier for sorting, extraction, rewording and drafts. Use the mid-tier or premium tier for texts published without editing. Premium is the tier for long analysis and demanding reasoning. The starter table gives the initial split. A blind test on your own tasks sets the final boundary.

Do you need a gateway to route?

No, because manual routing with a rule sheet captures most of the gain without any tool. Gateways bring a single key, automatic fallback and centralized rules. They make sense at high volume or in a team, once the statement has shown the need for them.

Disclosure. Savings vary widely with how usage is split, so your own statement is the reference. Price gaps between models can be checked on the official price lists cited in the sources. None of the providers mentioned pays us.

12Sources

1. OpenAI, official API price list
2. Anthropic, official Claude API price list
3. OpenRouter, official gateway documentation