AI Model Routing Explained: Auto, Custom, and When to Choose Yourself

TL;DR: Model routing decides which model answers a prompt. Custom routing keeps that choice with you; auto routing hands it to the platform. Reasoning effort and web search are separate switches that often move cost more than the model does, and routing is only a real choice when the models are already in one subscription.

Open a thread, type a question, and something has to decide which of the available models answers it. Routing is a question you can only ask once the labs have been consolidated into one subscription, which is why it barely comes up on a single-lab plan and why it arrives the week you replace two of them. On Perspective AI that decision is yours by default and can be handed off deliberately, which is the fork this page is about. Routing is the least discussed and most consequential control on a multi-model platform, and it belongs to model access because it determines which model you actually reach, whatever the catalogue says you have.

What Is AI Model Routing?

Model routing is the step that decides which model answers a prompt. Custom routing keeps the decision with you, per prompt or per thread. Auto routing hands it to the platform, which inspects the prompt and picks.

Custom Routing: You Pick, Per Prompt

Custom routing is a dropdown and a decision. You choose the model, you get that model, and the cost is the cost of that model's tokens. The reason to want it is that model behaviour is not fungible: labs differ in writing voice, instruction following, refusal patterns, code style, and how gracefully they handle a very long input. When one of those differences is the reason you opened the thread, no automatic selector can know that.

The practical pattern on a platform that supports mid-thread switching is to change model by task rather than by session. A long document goes to a large-context model for the analysis pass, the same thread moves to a strong writing model for the draft, and a fast cheap model handles the formatting and the follow-up questions. That workflow is only available where the thread survives the change, which is what switching models inside one conversation covers in detail.

Auto Routing: The Platform Picks

Auto routing replaces the dropdown with a classifier. The prompt is inspected, a model is chosen, and the answer comes back with no decision required from you. OpenRouter's Auto Router is the clearest published example on the developer side: its documentation states that the prompt "will be processed by a meta-model and routed to one of dozens of models", and that the response is priced "at the same rate as the routed model". No routing surcharge, and the pick is reported back.

The case for auto routing is not that it picks better than an expert. It is that most prompts do not deserve an expert decision. Across a working month, the accumulated seconds and attention spent choosing a model for routine turns is a real cost, and it is paid on every single message. Auto routing trades a small quality variance on unimportant turns for the elimination of that tax.

The case against is precision. A classifier reading a short prompt has very little to go on, and the prompts where the choice matters most are often the ones that look ordinary. The failure is quiet: you get a reasonable answer from a model you would not have chosen, and nothing tells you an alternative existed.

Two Switches That Move Cost More Than the Model Does

Perspective AI exposes three routing controls, and only one of them is the model picker. The other two change what a turn costs without changing which model answers it, which is why two identical prompts to the same model can produce different charges.

• Reasoning effort. Higher effort means the model generates additional internal tokens before it commits to an answer, and those tokens are billed like any others. On a reasoning-capable model this is frequently the largest single lever on the cost of a turn. It is worth spending on genuinely hard problems and close to wasted on retrieval, formatting, or summarisation.
• Web search. Turning search on adds a retrieval step and pushes fetched content into the context window as input tokens. It is the difference between an answer grounded in current sources and an answer from training data, and it costs accordingly. Leave it off for anything that does not depend on current facts.
• The model itself. The control everyone thinks of first, and often the third largest of the three in cost terms on any given turn.

A worked illustration makes the ranking concrete. Take one prompt and hold the model constant. At low reasoning effort with search off, you pay for your input plus a short answer. Turn effort up and the same model may generate several times that volume of internal tokens before it writes anything you can read, all billable. Turn search on as well and a set of fetched pages enters the context window as input on that turn and on every turn afterwards, because context is resent. Three configurations, one model, and a cost range wide enough that switching to a different model entirely would have been the smallest of the three changes.

The ordering is the useful part. Most people optimise model choice and never touch the other two, which means they are tuning the smallest of three dials. Whether any of this is visible to you at the moment of the decision depends on the platform, and that visibility is the same property that makes a metered credit balance readable in the first place.

When to Hand the Choice Off

A short decision rule, in order of how often it applies:

1. Hand it off for volume. Routine, repetitive, low-stakes turns where deciding costs more than deciding wrong.
2. Hand it off when you do not know the field. If you cannot name a reason to prefer one lab over another for this task, a classifier's guess is at least as good as yours.
3. Keep it when behaviour is the point. Voice, code style, refusal behaviour, or a specific model's handling of very long inputs.
4. Keep it when cost is the point. Deliberately routing simple work to cheap models is one of the more effective spending habits available, and an automatic router optimising for output quality will not do it as aggressively as you would.
5. Keep it for comparison. Running the same prompt across models on purpose is a routing decision that no auto mode will make for you.

Reading a Router's Receipts

Any routing system, automatic or not, owes you two facts per answer: which model produced it and what it cost. OpenRouter reports the routed model in the response payload and its activity log. A consumer platform should label the answer in the interface. If a platform routes silently and will not say where an answer came from, you cannot audit quality, you cannot audit cost, and you cannot tell a capable model from a cheap substitute. That is also the practical defence against the one genuinely dishonest pattern in this market, which is a service quietly serving a small open model under a frontier name.

The rate underneath all of it still comes from the labs, so before optimising routing at all it is worth knowing what the models you are routing between cost on their own plans.

Routing Only Exists Once Every Model Is in One Subscription

Routing between labs is not a feature you can buy from a lab. It appears the moment the models stop being separate subscriptions, which is the whole point of consolidating them: on Perspective AI all three controls sit in one thread on one $14.99/mo Starter plan, and the cost of the turn is shown before it is charged. You can run a thread on manual all week, flip a hard question to high reasoning effort for one turn, add search for a question that depends on this morning's facts, and never leave the conversation or reconcile a second invoice. Compared on how the routing decision is handled:

• Perspective AI: best for switching models and effort inside one thread, $14.99/mo.
• OpenRouter: best for programmatic routing from your own code, pay per token with the routed model reported back.
• Claude Pro: best when you have already decided one lab is the answer and routing is not a question you want, $20/mo.

Try routing a single thread across three models and watch which of the three controls actually moved the bill.

FAQ