Switchyard

ONE ENDPOINT.
EVERY MODEL.

THE GATEWAY

Put one API in front of every model you use. Route by cost, latency or eval score, fail over mid-stream, and keep the bill in one place.

99.98%

Requests served without a provider incident, last 90 days

Start routing

Routing production traffic for

NORTHWIND
halcyon
ferro/labs
Tidewater
Lumen.
PARCELPOST

Quickstart

SWAP ONE URL. KEEP YOUR CODE.

The gateway speaks the chat format you already send. Name a route, state a policy, and every response tells you which model answered and what it cost.

→Drop-in for the SDKs you use today
→Policies live on the route, not in your code
→Per-request cost and model in the response headers

POST /v1/chat

curl https://api.switchyard.example/v1/chat \
  -H "Authorization: Bearer $SWITCHYARD_KEY" \
  -d '{
    "route": "support-bot",
    "policy": ["cost", "eval>70"],
    "failover": true,
    "messages": [{"role": "user", "content": "Hi"}]
  }'

# x-switchyard-model: yard-4-mini
# x-switchyard-latency-ms: 92
# x-switchyard-cost-usd: 0.00041

Catalog · 9 models

BROWSE THE CATALOG

Nine models behind one key, priced per million tokens. Add them to a route and the gateway picks per request. Hover a plate to read its scores.

Read the routing docs

Routing policy

PRESS WHAT MATTERS

Four switches decide how the gateway picks a model for each request. The estimate below recalculates from the catalog as you press them.

Cheapest

Fastest

Top eval

Failover

Routing by cost, then eval score, with failover on.

$0.00 / 1M in

Pricing

PICK A TIER

Every tier runs the same gateway. What changes is how much of it you can hold open at once.

For side projects

HOBBY

$0 / month

1M tokens a month, any model
One route, two models
Community support

Most teams

TEAM

$180 / month

60M tokens, then metered
Unlimited routes and policies
Eval gates and failover
Spend caps per key

For regulated workloads

SCALE

Custom / annual

Private capacity and routing
Your VPC, your provider keys
SSO, audit export, SLA

Questions

Do I keep my own provider keys?

Bring them and we route through yours, or use ours and take one bill. Either way the keys are scoped per route and rotated from the dashboard.

What happens when a provider goes down?

Failover moves the request mid-stream to the next model in the route and marks the response, so your logs show which model actually answered.

How is the price per million calculated?

Provider cost plus a flat 4%, itemised per request in the log. Input and output tokens are priced separately, as they are by every provider. The policy console shows the blended input rate for the switches you have on.