unified inference gatewayunified inference gateway

Describe the job.We pick the model.Every request.

Tell us in plain English what your app does. We build a routing profile behind your key, then send routine work to cheap models and the hard stuff to frontier ones — so you get flagship results without flagship bills.

300+
models to route to
1
model id: auto
0
monthly fees
Base URL
https://api.airrouter.app/api/public/v1
API key
sk-gw-...

Create one in seconds after you sign up.

Model
auto

routes itself, or pin any model from the catalog.

Works anywhere an OpenAI-compatible endpoint is accepted.

200 OK · 412ms · charged $0.000184

one endpoint · every major lab

openai/gpt-4oanthropic/claude-sonnet-4google/gemini-2.5-prometa-llama/llama-3.3-70bmistralai/mistral-largedeepseek/deepseek-v3qwen/qwen-2.5-72bx-ai/grok-3cohere/command-r-plusamazon/nova-properplexity/sonarnvidia/nemotron-70bopenai/gpt-4oanthropic/claude-sonnet-4google/gemini-2.5-prometa-llama/llama-3.3-70bmistralai/mistral-largedeepseek/deepseek-v3qwen/qwen-2.5-72bx-ai/grok-3cohere/command-r-plusamazon/nova-properplexity/sonarnvidia/nemotron-70b
cost vs. capability

Frontier quality.A fraction of the cost.

Most of your bill goes on jobs that never needed your most expensive model. Match each job to the smallest model that still nails it — same output, a fraction of the spend.

The job to be doneInstead ofUse thisSaved / month*
See the full model catalog

*Estimated at the monthly volume stated on each row, using live catalog prices — figures move automatically when prices do.

Quality

No drop in output

These aren’t worse models — they’re the right size for the job. Frontier reasoning stays available for the work that actually needs it.

Effort

One line to switch

Same key, same OpenAI-compatible endpoint. Changing model is changing a string, so you can trial a swap in a minute and roll it back just as fast.

Proof

Real numbers, live

Rates here come straight from the live catalog, and your usage dashboard shows the per-request cost — so the saving is measured, not promised.

about a minute

How it works

step 01

Describe the job

In plain English: “summarise support tickets and draft replies”. No benchmark charts, no model shopping.

summarise support tickets and draft replies

step 02

We pick the models

You get a key with a routing profile — a cheap model for routine work, a mid model for tool use, a frontier model for the hard stuff. Edit any of them if you disagree.

  • Everydaygemini-flash
  • Balancedgpt-mini
  • Heavyclaude-sonnet
step 03

Send everything to auto

Change one base URL, use auto as the model id. Each request lands on the right tier automatically, and you only pay for what it actually needed.

{ "model": "auto" }

what you get

Infrastructure that getsout of the way.

Hundreds of models

Frontier and open-weight models behind a single OpenAI-compatible endpoint.

chatstreamingtoolsvision

Credits that refill themselves

Set a trigger and a top-up amount — your service never goes offline.

no subscriptionno invoices

Usage you can audit

Every request logged with tokens, latency, model and exact cost.

per keyper model

Keys you control

Issue as many as you need, revoke instantly. Stored hashed, never plaintext.

hashedinstant revoke

Move your mouse to seewhat you can automate.

One key, every frontier model. Ship your first request in about a minute.

  • automate: image generation
  • automate: voice generation
  • automate: transcription
  • automate: research agents
  • automate: content pipelines
Create an account

works everywhere

Paste three fields.It works everywhere.

Any tool that lets you set a custom OpenAI-compatible endpoint works with air router — no SDK, no migration.

  • Lovablepoint your AI provider at air router
  • Base44custom OpenAI endpoint in app settings
  • Boltswap the provider base URL
  • Cursoroverride OpenAI base URL in settings
  • Replitset the two env vars in your app
  • Windsurfcustom model provider
  • n8nOpenAI credential with custom base URL
  • Open WebUIadd as an OpenAI-compatible connection
  • LangChainbase_url on ChatOpenAI
  • …anything OpenAI-compatibleif it takes a base URL, it works

the three fields

1 · Base URL
https://airrouter.app/api/public/v1
2 · API Key
ar-...

create one in your dashboard

3 · Model
openai/gpt-4o-mini

or any model in the catalog

If a tool asks for an “OpenAI API key” and a “Base URL”, you’re done.