Cost, quality, latency
Bias the router. It re-decides every request.
Cost
Quality
Latency
RelayStart for free
The right model for every request.
Relay optimises each one across cost, quality and latency, then shows you the result on your own traffic before anything changes.
Costright-sized, never over-spec
Qualitynever routes below your floor
Latencybias to faster where it matters
30%+lower model spend
5%+accuracy gain
2lines to integrate
Coding agents and business appsMost routers only see your SDK traffic. Relay routes everything your company sends.
Two ways in.
Live swaps one base URL. Shadow writes no code at all — point your existing OpenTelemetry exporter at Relay and it reports what a different route would have done.
# before client = OpenAI(api_key=KEY) # after client = OpenAI(api_key=KEY, base_url="https://relay.oximy.com/v1") client.chat.completions.create( model="oximy/auto", messages=msgs)
Most work does not need the strongest model.
Teams wire one frontier model everywhere, then pay frontier prices to classify, extract and summarise. Relay reads each request and matches it to the model that fits, per turn, with a fallback so nothing fails.
How a route is chosen
Matched to the tier that fits.
Every request is judged against your quality floor before it moves. If a lighter tier clears it, that is where it belongs. If not, it stays where it was.
The routing surface
Everything that decides a route.
Auto or pinned
Choose per turn, or pin a model for one project.
Fallback model
A request always completes.
Routing reason
Engine, policy, or fallback — every row says which.
Route-only API
Ask which model. Make the call yourself.
Routing signals
Rules over model, tokens, cost, latency, source, error.
Token compression
De-duplicate repeated context across turns.
Every one of these is in the docs.
Endpoints, signals, transports, decision API.
Read the docsCustom routers
A router trained on your own traffic.
Off-the-shelf routers train on public benchmarks. Yours trains on the requests your company actually sends, including browser and coding-agent traffic that never passes through an SDK.
01
Create
Chat, code, or fully custom, scoped to one workload.
02
Pick candidates
Choose up to eight models the router may select between.
03
Train
It learns from your real requests, draft to ready.
04
Route
Swap one model id. Every call picks per turn.
Candidate models6 of 8 selected
Shadow mode
Prove it before you switch.
Run a candidate beside your primary on real paired traffic. Nothing changes until you say so.
What your last 10,000 requests would have costExample run
Spendlower is better−32.5%
Median latencylower is better−35.4%
Quality, judgedhigher is better+5.1 pts
Where the traffic wentshare of all requests
61%routed to a lighter model39%stayed where they were
Illustrative run on a 10,000 request sample. Your own numbers land in this same view once a shadow run clears, and nothing routes differently until you approve it.
LLM as a judge
A model grades the answers, not a person.
Nobody reads ten thousand answers by hand, so the cheaper model never gets a fair hearing. A third model does it instead, and scores each dimension separately.
It never learns which answer came from which model, and the order is shuffled every case, so position alone cannot decide the winner.
- Accuracy
Did it get the facts right
- Instructions met
Did it do what was asked
- Reasoning
Does the argument hold up
- Tone
Does it sound like your company
Transports
Pick how much of the path Relay touches.
| Mode | What Relay does | What changes for you |
|---|---|---|
| Observe | Ships traces only. | Keep calling providers directly. |
| Route only | Answers which model to use. | Relay never sees the completion. |
| Gateway | Decides and forwards. | One base URL. Your own keys. |
Pricing
See full pricingFree until it is doing real work.
Tokens
10M free every month
Then $0.15 per million routed tokens.
Routers
3 routers free
Then $99 per router per month.