01 / ROUTER
Policy-aware routing
Choose by task, region, compliance, latency budget, eval score or cost ceiling.
$ vanta init --production
One endpoint.
Every model.
Zero lock-in.
Route across frontier, private and specialist models through one production interface. Control quality, latency, cost and fallback without rewriting your application every quarter.
GLOBAL EDGE / 42MS P50TypeScriptPythoncURL
const result = await vanta.responses.create({
model: "auto",
policy: "support-prod",
input: "Summarize escalation risk",
fallback: true
})
frontier-aquality .95 / 820ms
specialist-rquality .92 / 410ms
private-7bquality .88 / 190ms
RESPONSE— ms
{
"status": "ready",
"model": "awaiting_request"
}MODEL —COST —EVAL —
router
policy
model
fallback
$ vanta benchmarks --live
Optimize the route,
not your loyalty.
Continuously evaluate models against your own tasks and route each request to the best available tradeoff.
MODELQUALITYLATENCYCOST / 1KROUTE SHARE
frontier-a0.95820ms$0.01834%
specialist-r0.92410ms$0.01141%
private-7b0.88190ms$0.00325%
02 / FALLBACK
Failure is a branch
Automatic retries and model fallback with deterministic limits and full traces.
03 / EVALS
Your benchmark, always on
Production samples become regression tests and routing signals—not dashboard decoration.
$ npm install @vanta/sdk