Starter
$49per month
- You keep
- 75% of verified savings
- Routing evidence
- Fleet-proven defaults
- Quality checks
- Fleet-wide
- Alerts
- Dashboard, daily digest
- Auto-accept
- No
- Rate limit
- 1,000 requests a minute
Two numbers. No surprises.
Our fee is two numbers, printed on the card: a monthly price, and the percentage of proven savings you keep. What you keep can only grow, because our fee is a fixed slice of savings we first prove on your own invoice. Your inference stays on your own provider accounts and bills: we never reroute it without your say-so, and we never bill for it. One URL to switch it on. The same URL to switch it off.
Pick your plan
Weekly checks, more models
Growth
$199per month
- You keep
- 80% of verified savings
- Routing evidence
- Your own traffic, weekly
- Quality checks
- Weekly, 5-point bar
- Alerts
- Everything in Starter, plus real-time email and webhook
- Auto-accept
- No
- Rate limit
- 10,000 requests a minute
Scale
$999per month
- You keep
- 85% of verified savings
- Routing evidence
- Your own traffic, weekly
- Quality checks
- Weekly, 3-point bar
- Alerts
- Everything in Growth, plus SMS and paging
- Auto-accept
- Yes
- Rate limit
- 100,000 requests a minute
The rest is our fee: a fixed slice of savings we prove against your own invoice, and only on routing changes you accepted. Our fee can only rise when the savings you keep rise more. A bigger plan buys you a bigger keep, your own traffic measured, and more ways to hear about a problem: every plan keeps all the alert channels of the plans below it. Never more protection: every safeguard ships on every plan.
Estimate your bill
Move the numbers to match your organization. Every figure derives from the assumptions shown, the same arithmetic your monthly statement uses, and your first month replaces assumptions with measurements.
Every plan gets every safeguard
- Screening before any request leaves, the full default detection set on every plan
- Routing only to providers you approved, through a routing table you own
- Budgets and caps that stop overruns before the invoice, per team, per user, per key
- A ledger that reconciles to your provider invoices
- Answers grounded in your own approved sources
- Audit-ready evidence export
A bigger plan never buys more protection. It buys better evidence, measured on your own traffic.
How the two numbers work
- The price pays for the product. On Growth and Scale, 80 cents of every dollar buys the model testing that measures cheaper candidates against your own traffic, at cost. Your back office shows every cent. Starter pays for infrastructure and carries no testing, so it costs what a governed gateway costs.
- You keep most of every saving, printed on the card. We recommend a cheaper model with evidence: measured quality parity, sample count, projected saving. You accept it onto your routing table, or you don't. No acceptance, no reroute, no fee. Each saving ties to the request that earned it, is priced at your own provider rates, and reconciles to your invoice.
- The cap you control is yours. Set hard token caps and dollar budgets per team, per user, per key, and the Gateway refuses past them. That bounds your provider spend. Our share needs no cap of its own: it is a fixed slice of savings you kept the larger slice of.
- Your provider account, your invoice. Your inference runs on your own provider account and is billed to you directly, at your rates, discounts included. We never bill for inference, mark it up, or touch the money.
Plans differ only in the evidence behind our recommendations: Starter routes on what we have proven across our fleet, Growth and Scale on what we measure weekly against your own traffic.
What a recommendation looks like
Your routing table shows each model you call, what serves it, and what we would change:
claude-opus → [claude-opus] recommended: [deepseek-v4 · 97% parity on your traffic · saves ~$4,100/mo]
Accept it and the route changes, the evidence is pinned to the decision, and we watch the quality continuously: a drop below the bar you accepted at alerts you and proposes a rollback. Decline it and nothing moves; the dashboard just keeps showing what the decline costs. Write your own route by hand and we claim nothing on it, and we watch nothing on it. That is the line the share pays for: not the click, the proof before it and the monitoring after it.
A worked example bill
A Growth customer spending $90,000 a month on inference accepts recommendations that save $11,000 in October.
| Line | Amount |
|---|---|
| Growth price | $199 |
| Our fee: 20% of $11,000 verified | $2,200 |
| Invoice from us | $2,399 |
| Savings kept by the customer | $8,800 |
| Provider invoice (paid direct) | $79,000 |
The dashboard shows the running share next to the running savings all month, so this bill is never news.
Why the bigger plans cost more
A quality comparison needs a fixed number of graded answers per candidate model, whatever your traffic: about 785 to prove a 5-point difference in win rate, about 2,180 to prove 3 points. Growth and Scale run those comparisons on your own requests, weekly, against progressively larger candidates, and 80 percent of the fee buys that testing at cost, priced from the first upstream's snapshot of 2026-10-27 (ADR-0002). Starter skips the testing and routes on fleet-wide evidence instead, which is why it costs a fraction and keeps the smaller slice: the price buys your keep up.
Objections, answered plainly
- "I accepted the recommendation, so I made the saving. Why pay you a fee?" The fee pays for what made accepting safe: the measured parity before the click and the continuous quality monitoring after it. Retype the same route by hand and we charge nothing, and you fly unwatched.
- "I do not trust a vendor's savings math." Every saving ties to the request that earned it, is priced at the rates we read from your own provider account, and reconciles to your invoice. The report carries the raw facts, so you can recompute every line.
- "What stops the bill from running away?" Arithmetic. Our fee is a fixed fraction of savings you verifiably kept the rest of, so every dollar on our invoice sits next to three or more that stayed with you, and the dashboard shows both running totals all month.
- "We will be locked in to your gateway." The Gateway speaks your clients' own API dialects, so your code never learns a new one. Leaving is the same change reversed: point the base URL back. Your policies and history export too.
- "A cheaper model will quietly get worse." Nothing moves without your acceptance, every acceptance is monitored against the bar it was accepted at, and a drop alerts you with a rollback proposal.
Savings you can check against your own invoice, a reroute that never happens without you, and an invoice that is a printed fraction of money you kept.
Start on one workload.
We admit customers in cohorts, in waitlist order. Don't miss the next one. When your slot comes up, you start on one real workload, on the plan you chose. The fee is on this page, and the share is billed only on savings verified against your own invoice. If we are not earning our fee, leaving is the same URL change, reversed.