The same tokens should not cost two prices.

The gap we sit in.

Frontier models are sold twice. There is the retail API price the labs publish, and there is the price of the same inference as cloud infrastructure, bought as committed capacity. The second is meaningfully lower, but it comes with procurement, quotas, and platform work that most teams sensibly refuse to take on for a line item.

Draftworks does that work once and passes the spread through: every model in the catalog at 20 to 25% under the provider's own price, behind the API contract you already use. Nothing about your prompts, your code, or the model changes. Only the bill does.

Why anyone should believe that.

A discount from an unknown middleman is not obviously good news. The industry's quiet failure mode is substitution: a smaller model, a quantized build, an aggressive cache, sold under the flagship's name. It is nearly undetectable case by case, which is exactly why it happens.

So the product is built to make trust unnecessary. Every response is signed over the model version and the exact bytes served, receipts are retrievable forever, and an open audit method compares our outputs against the vendor's own endpoint. If we ever served you something other than what the receipt claims, you would hold signed, permanent proof. The full argument is documented, because it is the reason the discount is safe to take.

What we deliberately are not.

Not a router with four hundred models: the catalog is nine, chosen because we can serve each one below list and verify it. Not a subscription: credits are prepaid dollars, spent by the token, refundable while unused. Not a platform that wants your data: content is hashed for the receipt and discarded, and the privacy default is stricter than the upstream's. The ambition is narrow on purpose. Serve the frontier, charge less for it, and prove it.