Inference
Get API key
OpenAI-compatible inference

One API for every model you actually ship.

One OpenAI-compatible API in front of every model. It fails over between routes when one stalls and reports exactly what it cost you. Keep your existing client and change one line.

Get an API key Read the docs

Free to start · No card required · Cancel anytime

—models in the catalogue
—upstream providers
99.9%routing target
The routing layer

Everything between your app and the model.

Requests in, answers out, with the reliability work handled before it reaches you.

Read the integration docs
  1. 01

    Failover, cooldown and retry

    Models with more than one route fail over automatically. A 5xx or timeout moves the request to the next; a route that rate-limits is parked on cooldown instead of being hammered. Your client sees one answer, not the switch.

  2. 02

    Drop-in OpenAI shape

    Chat completions, embeddings, images, speech and transcription all speak the OpenAI wire format. Point the SDK you already use at our base URL and keep your existing code.

  3. 03

    Keys and limits you control

    Create, rotate and disable keys from your console. Every key is scoped to your account and counts against one shared plan limit, with custom ceilings available on request.

  4. 04

    Usage and status with numbers

    Per-model requests, tokens and error rates, plus a searchable request log. The status page reports real route health and 24-hour success rate straight from the router, not a hand-set colour.

Live in about a minute.

No sales call, no onboarding queue.

  1. 01

    Create an account

    Register with an email and password. Your first API key is generated on the spot and shown once.

  2. 02

    Point your client at us

    Set the base URL to https://iceberg-inference.pages.dev/v1 and use any model from the catalogue as the model name.

  3. 03

    Watch it and scale

    Usage and logs fill in as you go. Move to Hobby or Pro when volume grows — or ask us for a quote at serious scale.

Reliability

Built for the messy middle of the market.

Frontier providers go down, rate-limit and retire models. Iceberg Inference tracks route health, parks failing routes on cooldown, and retries on another route when one is available.

  • Priority ordering across routes
  • Automatic retries with per-route cooldowns
  • Streaming preserved through failover
  • Honest error bodies that name the reason
Error response 
{
  "error": {
    "message": "The model is temporarily rate limited. Try again shortly.",
    "type": "rate_limit_error",
    "code": "rate_limit_exceeded"
  }
}

Questions, answered.

Still unsure? Email us — a person reads it.

Do I need to change my code?
If you already use the OpenAI SDK, change the base URL and the API key. The request and response bodies are the same shape. Anthropic-style clients can keep model as the only difference.
What counts against my limit?
Successful and failed inference requests both count toward your monthly request total, and prompt plus completion tokens count toward the token total. Listing models does not count.
What happens when I hit a limit?
Requests are answered with a clear 429 that names the limit and links to your plan. Nothing is charged overage without your say-so — you upgrade when you choose to.
Can I get a custom plan?
Yes. If you need committed volume, private routing or a service agreement, request a quote from the pricing page and we will reply with numbers.
Where do you run?
The router runs on our own infrastructure and forwards to providers over TLS. We keep request logs so you can debug and so we can run the service — you can read the policy in our privacy notice.

Start on the free plan.

Enough requests to build and demo. Upgrade the day real users show up.

Create account See pricing