“Cheapest LLM API” usually smells like a scam. Here’s one with receipts.
Most cheap-API pitches ask you to trust a discount you can’t inspect — resold quota, mystery keys. Ours is the opposite: keep your provider and your keys. The savings come from grading the work — the cheaper model only wins if it matches the quality — and every claim traces to a public run.
The smell test
You’re right to be suspicious.
Discounted tokens have to come from somewhere. When nobody will tell you where, it’s usually resold quota or somebody else’s keys — and you find out the day they get cut off.
A discount you can’t inspect.
- Mystery keys or resold quota sit between you and the model.
- Your traffic runs through someone else’s account.
- No way to check which model actually answered.
- The audit trail is “trust us.”
Cheap you can check.
- Keep your provider. Bring your own keys. One OpenAI-compatible endpoint — a faithful pass-through: same models, same outputs, read-only.
- The savings come from grading the work — a cheaper model only moves if it holds the quality — not from a mystery reseller.
- Every claim traces to a public run, re-gradable from the transcripts.
- $0 until you flip Pareto to production — a KEEP verdict pays us nothing.
And don’t expect your vendor to point out a cheaper grade: the cost-optimization docs at OpenAI, Anthropic and Google — 3 of 3 — never mention a competitor exists. Every vendor’s answer to “spend less” is “spend less with us.”
The receipt · Run №000 · 07.02.26
The cheapest 8/8 pass cost $0.0104 — 1/39th of the priciest.
Nine models. Identical prompt, one shot each, graded by an 8-point Playwright audit — not by taste. Five passed the same quality bar. The cheapest that did cost 1/39th of the priciest.
A five-way functional tie at a 39× price spread. That’s what “cheapest” means here: the lowest price that still passes the grade — not a discount, a verdict.
GPT 5.5 did ship the best-designed page — at 39× the price. When the frontier wins, the report says so.
Four models DNF’d on transport errors — failures in the pipes, not the models — receipts logged. Blind contests №001–005 are live on the public ledger ↗
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 the difference.
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