How we charge, and what we keep.
Plain-language summary; formal Terms of Service arrive with billing.
Launch terms (in effect now)
BYOK is free during launch. There is no platform fee on usage routed on your own provider keys while launch terms are in effect. Launch terms end when we announce pricing. You get 60 days’ notice, and organizations created during launch keep launch terms for 6 months after the announcement.
Managed keys are unchanged — provider cost at list plus 2%, drawn from a prepaid balance from the first request (we front the provider).
Adding a card lifts your fair-use caps and unlocks managed keys — it never adds a fee during launch. Orgs without a card stay free and capped: routed BYOK requests, telemetry retention, concurrent bake-offs, and published receipts each have a soft cap that a card on file raises. Caps are soft — exceeding one shows a notice and an “add a card” prompt; nothing is blocked mid-run.
How pricing works
One flat fee: 2% of what you run through ModelRig, valued at provider list prices — whether the API keys are ours or yours. Your negotiated provider discounts are yours to keep; our fee doesn't change based on which provider serves a request, which is part of why you can trust our routing. On your own provider keys, the first 1,000,000 requests each month are free, then 2%. Managed keys have no free tier — we front the provider's inference cost, so managed usage is provider cost + 2% from the first request. You also pay the standard Stripe transaction cost on a charge (ACH or card, at Stripe's published rates) — that is pass-through, not markup: our margin is the 2%. The probes, the registry data, the leaderboard, and the open-source tools are free, always. For comparison: the going aggregator markup is around 5% of model spend; ours is 2% flat at list prices — less than half the take rate, and the same meter charging it is the one proving what it saved you.
How billing works
Prepaid. You fund a balance; usage draws it down; when it's empty, requests stop with a typed error — your balance is itself a spending cap, so a runaway job can't surprise you. Balances never expire. Payment processing costs pass through at cost, itemized. Every statement is built from the same per-request telemetry you can see in your console — you can audit every line.
Refunds & disputes
Unused balance is always refundable — just ask. Consumed usage is not refundable: you can see exactly what ran, when, on which model, at what list price, in your own telemetry. That transparency is also how we respond to payment disputes: with the usage record.
Fair use
Managed keys run on our provider accounts, so provider content policies pass through to you; abuse gets accounts closed. Bring your own keys and your provider relationship is your own.
What we keep, and what you get for it
Two lanes, and you pick per route.
Pure router. Capture off, nothing retained. If a workload can't have a vendor holding payloads — data residency, ZDR, a compliance line you can't move — this lane is real, it is supported, and the routing is not a downgrade.
Optimization on — the recommended path, because it's the one that makes your bill fall. What's kept: the per-call telemetry you already see in your console, plus the review samples for bake-offs you create. Where: your organization, scoped by the database itself, not by our application code. How you leave: look at all of it in the console, delete it, export it, and take your routes with you — they were always YAML in your git.
What you get back: your first bake-off starts warm, on what other traffic has already proven about schemas shaped like yours, instead of cold. Evidence on your own routes, not a benchmark. Anonymized conformance stats — never payloads, never your schemas — are contributed by default, because that is what makes the next decision cheaper for everyone including you. It does not change your price: the same flat 2% whether you contribute or opt out to Pure router. We stopped discounting it because a price you can only get by giving something up is a bargain, and this is just how the product works.
We're not asking you to trust that we can't see your data. We're showing you every row of it, and letting you delete it.
The promises
Never a silent swap · your data never leaves without you · everything we hold, you can see and delete · one flat disclosed fee · your routes live in your git — leave anytime.