Switch to cheaper models. Keep quality.
Today you see model spend after the fact, and claimed savings are hard to check. The Gateway prices every request as it happens and stops overruns before the invoice. It tests cheaper models on a sample of your real traffic. When one clears your quality bar, it recommends the move with the evidence, and you decide.
- workload
- support-summaries
- current
- claude-sonnet-5
- candidate
- a cheaper open-weight model
- graded
- 785 answers against current
- result
- inside your 5-point quality bar
- action
- recommended, accepted by your admin
- evidence
- pinned to the decision, replayable
How it finds savings
- Sample. It replays a small share of real requests on cheaper models. Your users never see these replies.
- Grade. A judge model compares each candidate's answers with your current model's. It keeps grading until the difference is statistically clear.
- Recommend. When a cheaper model clears your bar, a recommendation appears next to your route, with the evidence behind it: parity on your own traffic, and what it saves. You accept it onto your routing table, or you don't. Nothing moves without you.
- Watch. An accepted route is monitored against the bar you accepted at. A drop below it alerts you and proposes a rollback. New models and prices arrive every month, so the search never stops.
A recommendation applies only to the kind of work its evidence covers. A model that proved itself on summaries earns no claim on your code review, and a request the Gateway cannot classify with confidence serves exactly the model you chose, with no savings claim.
A smaller bill, with proof
- No surprise invoices. Budgets warn, throttle, switch to a cheaper model, then refuse, before you overrun.
- Savings per request. Each request records what it saved against the model your admin chose for it. That choice is the baseline, and it is yours, so our own optimizations can never quietly become the new zero.
- Matches your invoice. Savings and spend reconcile to the provider invoice you actually pay, not a price table we maintain, with every difference explained. Spend reports by company, group, and person, and chargeback falls out of the ledger.
- Quality stays put. An accepted route never drops below the quality bar you accepted it at, and a drop proposes a rollback.
- Pin anything, pay for nothing. Pin a workload and it serves exactly the model you chose. We claim no savings on pinned traffic, and the dashboard shows what the pin forgoes, as plain arithmetic.
Most of your bill is a share of what we save you, checked against your invoices. Save nothing, and that part is zero.
Objections
- "I do not trust a vendor's savings math." Every saving ties to its request and reconciles to the invoice. Export it and check it yourself against your own provider bill.
- "A cheaper model will quietly get worse." A recommendation appears only after a judge model clears your quality bar on graded answers, nothing moves until you accept it, and the accepted route is monitored against that same bar.
- "Why pay a subscription if savings cover it?" It pays for the search, whether or not a cheaper model turns up that month. 80 cents of each dollar runs it, and you see every cent.
- "Another hop will slow my feature." Pricing and grading run alongside the request, not inside its critical path. Overhead is a published target, under 60 ms added, 40 ms to first token.
Check our math yourself
Export the savings for any period.
- Reconcile them against the ledger and your provider invoice.
- Open the evidence behind any accepted recommendation, and replay its trial run.
See your savings before you pay.
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 your plan from day one. You keep a savings report you can check against your own invoices. If we are not earning our fee, leaving is the same URL change, reversed.