FAQ
The questions people actually ask about Coco Router.
What is Coco Router, in one sentence?
A single OpenAI-compatible endpoint and API key that automatically routes each request to the cheapest model capable of handling it — cheap, fast models for simple prompts, strong models only when a task genuinely needs one. See Core concepts.
How much does it actually save?
Compared with sending everything to a frontier model, typically 40–80%, depending on your traffic mix. Trivial questions run on a small model for fractions of a cent while hard tasks still get frontier-class quality. See Auto-Switch.
Do I have to change my code to use it?
No. Point any OpenAI-compatible client at the gateway base URL and use a Coco
Router key. Use an auto-* key and you don't even change the model name your app
sends — it's routed regardless. See Integrating an application.
What's an "auto key"?
A virtual key whose alias starts with auto (e.g. auto-myapp). Every chat
request through it is routed by complexity, and the model name the caller sends
is ignored. This is what makes Coco Router drop-in for third-party tools that
only let you pick a model from a fixed dropdown — pick anything, it's routed
anyway. See Virtual keys & auto keys.
How does it decide which model to use?
A small, cheap classifier model reads each prompt and sorts it into one of four tiers — SIMPLE, MEDIUM, COMPLEX, or REASONING — by actual cognitive difficulty, not keywords. Each tier maps to a priority list of models. Classification costs roughly $0.00002 and adds ~200–400 ms. See Auto-Switch.
Can I change which models each tier uses?
Yes, from the Auto-Switch Control Panel. Edit the model pool for a tier and click Save changes — it applies to the next request, with no restart or redeploy.
How do I see which model answered a request?
Two ways. The response body carries an ai_router object with the tier and the
real model; and the response includes x-ai-router-tier / x-ai-router-model
headers. Browsers can't read custom cross-origin headers, so from front-end code
use the body field. The AI Console shows every decision live.
Which models and providers are available?
Over 1,300 model entries across 18+ providers, including the full OpenRouter catalog and Amazon Bedrock. Some providers are active now; others are wired and go live the moment their key is added. See Models & providers.
Does routing affect audio transcription or embeddings?
No. Auto-routing applies only to chat completions. Audio transcription, embeddings, image generation, and other non-chat endpoints pass straight through to their real models — even on an auto key. So one key handles chat, voice, and RAG. See Audio, embeddings & non-chat endpoints.
What happens if a model is down or rate-limited?
Each tier holds a priority list. If the first model errors or rate-limits, the router fails fast and falls to the next one — usually within a few seconds. This is what makes "free model first, cheap paid model as backup" safe in production.
Are free models reliable?
Free models cost $0 but are heavily rate-limited under real traffic. The recommended pattern, and the shipped default for the SIMPLE tier, is free-first with a cheap paid fallback so you get $0 when a free model is available and seamless failover when it isn't. See Auto-Switch Control Panel.
How is billing tracked?
Every request is attributed to the virtual key that made it, with its cost. Give each application its own key and the Dashboard gives you per-app cost reporting for free, plus per-model breakdowns.
Can I stop one app from spending too much?
Yes. Set a Max Budget on its key — a hard lifetime cap that stops the key when reached — and you can block or delete the key from the Dashboard at any time, without affecting any other app. See Security.
Can external AI tools or agents call it?
Anything OpenAI-compatible can. Give it an auto key and point it at the gateway. See Integrating an application.
Is there a public sign-up?
No. Credentials are provisioned by the Cocolevio team. See cocolevio.com/contact.