# Smart routing and auto: same answers, faster and for less

> CArouter can now pick the model for each request, and decide whether it thinks first. On our tests, auto answered better than any single model, and smart mode cut a model's bill by 50%.

- Published: 2026-11-14
- Author: Flavien Berwick
- Category: Product
- Web page: https://carouter.ai/blog/smart-routing-and-auto
- In French: https://carouter.ai/fr/blog/smart-routing-and-auto.md

![Product: smart routing and auto, same answers, faster and for less. Auto, answers that pass: 80% for the best single model, 86% for auto. Smart mode with GLM-5.3 named: average wait, 22 then 17 seconds; cost of the requests, CAD 13.18 then 6.54. 328 graded requests in English and French, nine kinds of work, served from Canada.](https://carouter.ai/blog/smart-routing-and-auto/illustration.png)

Most of what a team sends to an AI model is simple: a thank-you, a translation, a short question. Today those requests get the same treatment as the hard ones. They go to the team's largest model, which thinks at length before it answers, at full price.

CArouter can now choose, for each request, the model most likely to answer it well, and decide whether that model needs to think first. On the requests we graded, auto gave more good answers than any single model: 85.7% against 79.9%. With :smart on a model you already use, the share of good answers held within two points and the bill fell by up to 50%.

## The short version

- **What it does.** For each request, CArouter picks the model most likely to answer it well, takes the least expensive when several are as good, and skips the model's thinking time when the request does not need it.
- **How you use it.** Your application changes one word: it writes auto where it names a model today, or it keeps the model's name and adds :smart.
- **What we measured.** On 328 graded requests, auto passed 85.7% of answers, where the best of three single models passed 79.9%. With :smart on GLM-5.3, the most expensive of the three, 78.0% of answers passed against 79.6%, the requests cost CAD 6.54 instead of 13.18, and an answer took 17 seconds instead of 22.
- **What stays the same.** Your requests stay in Canada unless your key says otherwise, you pay the price of the model that answered, and every response names that model.
- **Availability.** Both run on our own workspaces first. We will open them to every workspace next.

## The problem: one model for every request

A team usually picks one model and sends it everything. The model is chosen for the hardest work the team does, so it is large, it thinks before it answers, and it is priced to match.

Now look at three requests the same team sends in one morning.

- **"Thanks, that worked!"** Nothing to think about.
- **"Translate this paragraph into French."** Any good model does it in two seconds.
- **"List every amount and deadline in this contract."** This one deserves the large model, and its full attention.

With one model for everything, the first two wait as long and cost as much as the third. Across a day of requests, that is most of the bill and most of the waiting.

## What CArouter does now

CArouter reads each request before sending it anywhere, and makes two decisions that your team used to make once, for every request at the same time.

**Which model answers.** Each request goes to a model we have measured to handle that kind of work well. When several do, the smaller, faster, cheaper one answers.

**Whether the model thinks first.** Many models can work through a problem in a private draft before they write their answer. You wait for that draft and you pay for it. CArouter turns it off when a direct answer is just as good, and leaves it on when the work calls for it.

The same three requests now go three different ways. The thank-you is answered at once, with no thinking. The translation gets a direct answer in a couple of seconds. The contract goes to the large model, thinking included.

## Two ways to turn it on

Both are one word in the request your application already sends. Nothing else changes in your code.

*The same three requests, sent three ways: today, with auto, and with :smart.*

Today all three requests go to one model, which thinks before every answer. With auto or :smart, the thank-you and the translation are answered directly in a couple of seconds, at a fraction of the price, and the contract still gets a large model's full attention.

### Let CArouter choose: write auto

Where your application names a model, it writes auto. CArouter then picks the model and decides on the thinking, request by request. For each request, the standard is the model most likely to answer that kind of work well, among those your key can reach. When several are as good, the least expensive and fastest one answers.

This is the option for a team that wants the best answer to each request and has no attachment to a particular model.

### Keep your model as the standard: add :smart

If your team has settled on a model, it keeps that model's name and adds :smart after it. Your model still answers every request that needs it. A request goes to another model only when that model is expected to answer it as well, cost no more and take no longer.

This is the option for a team that has tested a model, trusts it, and wants to stop paying for it on the easy requests.

### What you send, and what comes back

A request that lets CArouter choose looks like any other. Only the model field differs.

```
curl https://carouter.ai/v1/chat/completions \
  -H "Authorization: Bearer $CAROUTER_API_KEY" \
  -d '{"model": "auto", "messages": [{"role": "user", "content": "Translate this paragraph into French: ..."}]}'
```

The response carries two lines your team can log: the model that answered, and what CArouter read in the request to choose it. Here a translation, judged easy, answered with thinking off.

```
X-CArouter-Model:   qwen3.6-35b-a3b
X-CArouter-Routing: {"mode":"auto","kind":"translation","difficulty":0.6,"effort":"none"}
```

## How CArouter decides

*One request after another: CArouter reads the request, weighs the candidate models against a quality bar, picks the least expensive one that clears it, and names it in the response.*

- **It reads the request.** A small model on our own hardware in Canada looks at the request for a few tenths of a second: what kind of work it is, how hard, and whether the answer will be long. A thank-you is recognised with no model at all.
- **It knows how each model does on that kind of work.** We graded each model on hundreds of requests, by kind of work. That record is refreshed every ten minutes with current prices and measured speeds.
- **It only answers when quality holds.** A model answers a request only when its record on that kind of work is as good as the standard's, with a safety margin for what we are less sure of. The margin is set so that the share of good answers stays within two points.
- **It has a fallback.** If the chosen model fails or returns a broken answer, the standard for that request answers instead, before you receive a single word.

If CArouter cannot read a request for any reason, the request is served exactly as it was written. Choosing a model is a saving, and it never stands in the way of an answer.

## What we measured

We wrote 328 requests of the kind our customers send, half in English and half in French: code, math, research, business writing, extraction, summaries, translation, everyday questions and agent steps. Each answer is graded by a test, an exact result or a rubric.

Three open-weight models served from Canada answered all of them: Qwen3.8 Flash Next, GLM-5.3 and Qwen3.6 35B A3B. Qwen3.6 35B A3B answered them twice, with its thinking on and with its thinking off.

Every figure below is measured out of sample: to score a request, we choose every setting on the other requests, then see how that request fares.

![Answers that pass against the cost of the 328 requests. Qwen3.8 Flash Next: 79.9% for CAD 4.91. GLM-5.3: 79.6% for CAD 13.18. Qwen3.6 35B A3B with thinking off: 76.8% for CAD 0.57. Qwen3.6 35B A3B: 75.9% for CAD 3.16. Auto: 85.7% for CAD 3.73. Smart mode with GLM-5.3 named: 78.0% for CAD 6.54. Choosing the best model for each request after the fact: 91.8%.](https://carouter.ai/blog/smart-routing-and-auto/01-quality-against-cost.png)

### One model for everything

Taken one at a time, the three models are close: they pass between 75.9% and 79.9% of the requests. What separates them is the price, from CAD 3.16 to 13.18 for the 328 requests, and the wait, from 17 to 69 seconds on average.

They are close on the total and different in the detail. Qwen3.8 Flash Next passes 92% of the math requests and 44% of the agent steps. Qwen3.6 35B A3B passes 66% of the agent steps and 67% of the math requests. No model is the best at everything.

### Auto

Auto sends each request to the model most likely to answer that kind of work well. It passed 85.7% of the answers, 5.8 points more than the best single model: out of 100 requests, 6 more good answers.

The gap is larger than the noise of a set this size: when we resample the requests, it stays between 2.4 and 9.4 points.

It used every model: 52% of the requests went to Qwen3.8 Flash Next, 31% to Qwen3.6 35B A3B with thinking off, 11% to GLM-5.3, and the rest to Qwen3.6 35B A3B with thinking on.

![Average wait for a complete answer on the 328 requests. Qwen3.6 35B A3B: 69 seconds. GLM-5.3: 22. Qwen3.8 Flash Next: 17. Smart mode with GLM-5.3 named: 17. Auto: 9.8. Qwen3.6 35B A3B with thinking off: 6.4.](https://carouter.ai/blog/smart-routing-and-auto/02-average-wait.png)

An answer took 9.8 seconds on average and the requests cost CAD 3.73. The best single model, Qwen3.8 Flash Next, cost CAD 4.91 and took 17 seconds, and GLM-5.3 more on both counts. Qwen3.6 35B A3B alone costs CAD 3.16 and answers in 69 seconds, with 10 fewer good answers out of 100.

### Smart mode

Smart mode keeps the model you named as the standard and serves a request from another model only when that model is expected to answer as well, cost no more and take no longer. With GLM-5.3, the most expensive of the three, it passed 78.0% where GLM-5.3 passed 79.6%. The requests cost CAD 6.54 instead of 13.18, and an answer took 17 seconds instead of 22.

The saving depends on the model you name. With Qwen3.6 35B A3B: 79.6% against 75.9%, CAD 2.86 instead of 3.16, and 37 seconds instead of 69. With Qwen3.8 Flash Next, no other model was expected to answer as well for no more money and no more wait, so smart mode served every request from it.

In every case the share of good answers stayed within the two points smart mode is set to keep. Our research page shows every figure.

## Thinking helps on some work, and costs on all of it

To learn when thinking is worth its wait, we asked Qwen3.6 35B A3B all 328 requests twice, with thinking on and with thinking off. With thinking, it passed 75.9% of them and an answer took 69 seconds on average. Without, 76.8% and 6.4 seconds.

The average hides where thinking matters. On business writing, 27 of 28 answers passed with thinking and 21 without. On code, 46 of 66 passed with thinking and 56 without, and an answer took 127 seconds longer. On code and math, 30 of the 130 answers were still thinking when they reached the length limit, and count as failed.

![Qwen3.6 35B A3B on the 328 requests. Always thinking: 75.9% of answers pass, 69 seconds on average. Never thinking: 76.8%, 6.4 seconds. Thinking decided for each request: 79.3%, 14 seconds.](https://carouter.ai/blog/smart-routing-and-auto/03-thinking.png)

Always thinking makes every request wait and costs CAD 3.16 for the 328 requests. Never thinking costs CAD 0.57 and loses the work that needs it. Deciding for each request skipped the thinking on 74% of them and passed 79.3%, more than either, for CAD 0.88 and 14 seconds on average. Auto goes further, because it can also change model.

## What we tried that did not work

We hoped to catch a weak answer from its first words. A model reports how sure it is of each word it writes, and a hesitant start could have told us to switch to a stronger model before you saw anything.

On our requests that signal was close to a coin toss. We built the check, measured it, and left it switched off. A safeguard that does nothing has no place in a product, and you should know which ones we tested.

## What stays the same

- **Your data stays in Canada.** A model that CArouter chooses for you is served from Canada by a Canadian-owned provider. A request crosses a border only for an API key that allows foreign providers.
- **The bill is the price of the model that answered.** With :smart, that is never more than the model you named.
- **Every choice is visible.** Each response names the model that answered and the reason, in a header your team can log.
- **Your prompts are not kept.** Reading a request leaves two labels behind: the kind of work and its difficulty.
- **Nothing changes unless you ask.** A request that names a model and adds nothing is served by that model, as before.

## New models, new prices, every week

New models arrive every month, and prices and speeds move every week. A list of rules written by hand is out of date the day after it is written.

CArouter's record is rebuilt every ten minutes from the models, prices and speeds in service. A new model starts taking requests on your behalf once it has answered our graded requests, and until then it is served only when you name it.

## What this does not prove

- This is one set of requests, written and graded by us. A different set could move the figures.
- Every request is scored with settings chosen on the other requests. An earlier version of this article used 137 requests kept aside, which we read several times as the work changed.
- An answer that ended in an error was asked a second time, and counts as failed if it failed again.
- Asked the same 328 requests a second time, Qwen3.6 35B A3B changed result on 45 of them. A difference of a point or two between two rows is within that noise.
- The choices were replayed on answers already graded. In production a request also waits a few tenths of a second to be read.
- Auto relies on how each model did on each kind of work. A model that changes, or work unlike our nine kinds, needs new measurements.

## Built on published research

None of these ideas is ours alone. Sending easy requests to a smaller model, describing a new model by its results on a fixed set of requests, and setting a quality margin on data kept apart all come from published research. Our research page lists each paper with what it found.

[See how it works, with the measurements and the references.](https://carouter.ai/research)

[An earlier article on the small model that reads each request: Kev is available on CArouter.](https://carouter.ai/blog/kev-available-on-carouter)

The right model for a request is the least expensive one that answers it well. Finding it is now CArouter's job.

[Compare available models and providers.](https://carouter.ai/docs) [Read the technical routing rules.](https://carouter.ai/llms-full.txt)

---

More articles: [CArouter Blog](https://carouter.ai/blog.md). What CArouter is and how to call it: [llms.txt](https://carouter.ai/llms.txt).
