Platform

Model verification

Each model on Zurelay is served over several routes, run by independent sellers. Every route is tested before it serves requests and every day after, so the model you name is the model that answers.

Why prices are lower

Model capacity is often bought ahead, through prepaid credits, committed-spend contracts and volume tiers, and not all of it gets used. Sellers offer capacity like that below list price. Zurelay tests it, pools several routes per model and sends each request to one that is available, verified and well priced. More on the Sellers page.

Before a route serves requests

Identity. The model is asked who it is under different system prompts and personas. Naming another product, or dropping the instructions it was given, fails.
Knowledge. Questions about dated events with known answers catch an older or cheaper model sold under a newer name.
Reasoning. Problems with one exact answer catch a smaller model swapped in for a larger one.
Response details. The model name and token counts the route reports are compared with what was sent.
Images. Requested sizes and aspect ratios are measured on what comes back.
Tools and streaming. Tool calls in OpenAI and Anthropic formats, and streaming, have to work as documented.

The exact tests aren’t published, so they can’t be gamed.

While it serves

When	What happens
Every 10 minutes	Each chat model gets a test request of ours, on top of live traffic. A route that keeps failing is moved to the back of the line.
Every day	Knowledge and identity checks run again: a seller can change what runs behind a route at any time.
Every request	An error page sent as an answer, or an image at the wrong size, counts as a failed attempt. The request moves to another route and you aren’t billed for it.
When a route fails a check	It’s taken out of rotation until it passes again. The model’s other routes keep serving it.

What it means for you

You’re billed for the tokens you send and receive, never for instructions a route adds behind the scenes.
Live health for every model is on the status page, and Reliability explains failover and timeouts.
If an answer doesn’t look like the model you asked for, open a request on the Support page with the response’s x-request-id. We check the route that served it.

What the checks cover

They test quality and reliability: that you get the model you asked for, working as it should. They aren’t audits of a seller’s security or data handling, and requests aren’t served under a model maker’s own terms for its customers. Our Terms (sections 3.3 to 3.6) explain how requests are served.

Selling through Zurelay

Have capacity to spare? Apply on the Sellers page. A person reviews every application, and every route passes the same checks before any customer traffic reaches it.

Questions, or something missing? Ask our support team and we’ll answer by email, usually within a minute.