The AI gateway that shows you what the other model would have said.
One OpenAI-compatible endpoint. Your own keys, faithful pass-through — same models, same outputs, read-only — while our model grades every workload in shadow. $0 until you flip.
The category, corrected
Gateways move traffic. We grade it.
Any gateway can put your request on a model. Almost none can show you, with evidence, whether that was the right model — or where the same quality was available for less.
Routers pick a model and ask for trust. We show you both answers — and where the savings came from.
Every comparison is re-gradable from the transcripts. If you doubt a verdict, re-run the grade yourself — the receipts are yours.
Why the umpire has to be independent: the cost-optimization guides from OpenAI, Anthropic and Google — 3 out of 3 — never mention that a competitor exists. Every vendor’s answer to “spend less” is “spend less with us.”
One endpoint
Keep your provider. Add a witness.
One OpenAI-compatible endpoint in front of the keys you already have. Nothing about your stack, your models, or your outputs changes — you just see where you could get the same quality for less.
↓
shadow report — async, off the production path
Point your SDK’s base URL at it and you’re done. You keep your provider — and flip when ready, workload by workload.
Bring your own OpenAI and Anthropic keys. You pay provider prices on your own keys — the pass-through itself costs $0. Keys encrypted, never logged.
Same models, same outputs, read-only. Your users see exactly what your provider returns — the gateway never touches an answer.
Proxy, shadow and reports are free. We charge only when you flip Pareto — our model — to production, at ≈ half frontier list. A KEEP verdict pays us $0.
Receipt: Run №000 · 07.02.26 · blind ledger №001–005
Why grading pays
Same passing grade. 39× the price.
Run №000 · 07.02.26 · nine models, one shot each, an 8-point Playwright audit. Five passed 8/8. What the same passing grade cost:
39 times the price spread for the same passing grade — the gap a gateway should be showing you on every workload.
And sometimes the priciest answer is worth it: GPT 5.5 shipped the best-designed page in the run. That’s what a KEEP verdict is for.
Four models DNF’d on transport errors — failures in the pipes, not the models: Fable 5, Opus 4.8, Kimi K2.5, MiniMax M2.7. №000 is our open run; the blind contests are on the public ledger (№001–005).
Receipts
Same prompt. Both answers. You judge.
Pick a prompt. Two answers come back — one from the frontier model, one from the value model. You decide whether the grade holds.
Real runs through our own gateway, July 7 2026 · answers untouched · costs are what the meter read
The mechanism
Proof, not promises.
Point your OpenAI and Anthropic keys at our pass-through proxy. Same models, same outputs, read-only. You keep your provider.
It re-runs a sample of your real requests async, off the production path. Users never see it.
Quality and price per workload, re-gradable from the transcripts. Nothing moves without your sign-off.
Pareto wasn’t ready to enter Run №000 — once it ships, it enters every contest under the same audit.
Live July 15
Send us your AI bill.
We’ll show you where the same quality costs less.
We’ll write when unbiased.ai opens on July 15. First shadow runs go to the list in order.
Join before July 15 → $50 in credits when we open · keys encrypted, never logged · free until you flip Pareto to production
Prefer email? Send your bill to hello@unbiased.ai