Five SLA questions to ask any API gateway before you top up

Last updated · Written by Nova

Nov 5, 2026 · #deepseek#api#reliability

"99.9% uptime" on a landing page is a marketing number until you can measure it. What you actually need to know is how a gateway behaves when something breaks — because something always breaks. Five questions cover it.

1. What happens to an in-flight request when the upstream hiccups?

The useful answer is specific: whether a failed upstream call is retried, how many times, and whether a retry is billed to you. A gateway that silently retries three times and bills all three is more expensive than its rate card suggests.

2. How do I know which model build I am getting?

Families keep stable names; builds change. A gateway that will not publish its upstream build is one you cannot hold accountable when behaviour shifts without any code change on your side.

3. Is there a status page, and is it honest?

A status page that has never shown a yellow day is either new or decorative. You want incident history: what broke, when, how long, and what was done. The existence of history is worth more than a green badge.

4. What does the error body look like when billing fails?

Trigger a 402 on purpose before you commit money. A gateway that returns a generic 500 for a billing problem is one you cannot debug at 2am. The error text is the difference between a five-minute fix and an evening of guessing.

5. How does the balance and refund flow actually work?

Ask three specifics: is the balance prepaid, does it expire, and how is an unused balance returned. Then read the terms page and check the answers match — a mismatch between marketing copy and terms is the loudest warning sign there is.

What to test before the first real payment

  1. Send the smallest possible top-up and confirm the credit arrives.
  2. Trigger a deliberate failure (empty balance, wrong model name) and read the error bodies.
  3. Check usage reporting after a day of real traffic against your own token counts.
  4. Only then consider a serious balance.

Our own answers to these five are public: how the route is built, the endpoint reference, and the terms — including refund handling, stated plainly.

Questions people actually ask

What does SLA actually mean for an API gateway?

For most small gateways it means no formal SLA at all. What matters is measurable behaviour: retries, incident history, honest status pages, and billing transparency.

Why should I care about the upstream model build?

Builds change behaviour without any change on your side. A gateway that publishes its build is one you can hold accountable.

How do I test a gateway's error handling?

Trigger a 402 on purpose and read the error body. Vague errors at billing time predict vague errors everywhere.

Is a status page with no incidents a good sign?

Not necessarily. It can mean new, or decorative. Incident history with resolutions is worth more than a permanent green badge.

What is the safest first payment?

The smallest top-up the service allows, followed by deliberate failure tests, before any meaningful balance.