Quick setup
Open the kRouter dashboard at http://localhost:20128
Go to Providers → click Deepgram
Paste your API key, click Save
Test the connection
Use in your IDE
Point any OpenAI-compatible client at kRouter using your local API key.
# Endpoint
http://localhost:20128/v1
# API key (from Dashboard → API Keys)
sk-krouter-XXXX
# Model
dg/<model-id>How Deepgram billing works with kRouter
Deepgram is pay-per-token. You bring your own API key and kRouter routes to it, so you keep your own rate limits and billing relationship. kRouter authenticates it via apikey. 3 models are routable as deepgram/<model>.
Through kRouter, Deepgram serves speech-to-text, exposed on /v1/audio/transcriptions, image understanding (OCR / vision-to-text), text-to-speech, exposed on /v1/audio/speech. Requests reach it through the same local OpenAI-compatible endpoint as every other provider, so switching to or from Deepgram is a one-line model change in your client rather than a rewrite.
$200 free credit on signup (no card required). Aura-1: $0.015/1k chars, Aura-2: $0.030/1k chars (Pay-As-You-Go).
Deepgram models & pricing
These are the model ids kRouter routes to. This provider does not publish per-token rates.
| Model | Route as |
|---|---|
| Nova 3 | deepgram/nova-3 |
| Nova 2 | deepgram/nova-2 |
| Whisper Large | deepgram/whisper-large |
Deepgram links
Deepgram FAQ
Is Deepgram free through kRouter?
Deepgram is pay-per-token: you bring your own API key and are billed by Deepgram at their rates. kRouter adds nothing on top and is free and MIT-licensed.
How do I use Deepgram with Claude Code or Cursor?
Connect Deepgram in the kRouter dashboard, then point your tool at kRouter's local endpoint (http://localhost:20128/v1) with a kRouter API key. Any OpenAI-compatible client works, and models are addressed as deepgram/<model>.
Which Deepgram models can kRouter route to?
3 models, including Nova 3, Nova 2, Whisper Large. Every one is addressed as deepgram/<model> from any client.
What happens when Deepgram is rate limited?
kRouter detects the limit, parks that account until its real reset time, and fails over to your next connected account or provider — so the request still completes instead of erroring.