Most prompts
do not need
the frontier.

SwitchRoute sends each request to the cheapest model that can still do the job.

For teams paying OpenAI, Anthropic or Google per token.

Start free for 30 days How it works ↓

base_url = "https://switchroute.dev/v1"  # was api.openai.com

Same OpenAI-style requests. Change the base URL and key.

How does it work

A prompt landing on a tier

  • T1 Easy
    • LunaOpenAI
    • FlashGoogle
    • HaikuAnthropic
  • T2 Everyday
    • TerraOpenAI
    • SonnetAnthropic
    • Gemini 3.8 FlashGoogle
  • T3 Hard
    • SolOpenAI
    • OpusAnthropic

The cheapest tier that can still do the job.

You choose the model on each tier. If even the easy work is hard, put Sol on Easy instead of Luna. Those requests then cost Sol’s price.

What you do

  1. Add your provider key. OpenAI, Anthropic or Google. We store it securely and use it only for your requests. The model company bills you directly.
  2. Choose a model for each tier. Easy, Everyday and Hard. You decide what runs where.
  3. Create a SwitchRoute key. It starts with sr_live_ and is not the same as your provider key.
  4. Change the base URL in your app or tool. Use https://switchroute.dev/v1 and the new key. Requests stay OpenAI-compatible.

Four steps, no card.

The console walks you through each one. The first 30 days are free.

Start free for 30 days

Already run your own AI? Plug it in.

Your own model can be one of the choices.

If your company runs its own AI model, or you use a service we do not list, add its address in the console. It then shows up next to the other models, and you can give it any of the three tiers.

Easy jobs can go to your own model, and the hard ones to a big provider. If your model is down, the request goes to a backup, so nothing stops.

Observability and data protection

Easy · Luna · $0.04

Label this review as positive

You can always see the route and the cost.

The words are optional. We keep them for 30 days unless you change that. Turn storage off and we do not keep the prompt or the reply.

When the days end, that request is deleted. Daily totals and billing records stay. Privacy notice.

Price

of the difference
20%
a request
$0.005

Free for 30 days.

The trial ends after 30 days or $10 of waived fees, whichever comes first. Then 20% of the measured difference plus about $0.005 a request. If the difference is zero, the share is zero. The model companies bill you directly.

We never take more than half of what a request saves you. If a request saves nothing, for example because you already send it to the cheapest model, you pay only the small per-request charge. A request never offsets another request.

ExampleYou spend $1,000 a month sending everything to Opus. In our own analysis routing saved 34% on average, so $340. We take 20% of that ($68) plus the per-request charge, and never more than $170 in total. You keep the rest.

Start free for 30 days

What if it fails?

20% 0

Easy

Hard

If we mess up, your routing fee for that request is waived.

If your next message says the answer was wrong, we waive that fee and retry on a stronger model.

A classifier also judges whether the task was completed. If it was not, we waive that fee.

The waived fee is the 20% and the per-request charge. The model company still bills you for the call.

See what your prompts actually need.

Start free for 30 days