How to choose a model gateway: what to look at

There are many model gateways, and they differ less in the models they carry than in what happens when something goes wrong. Below are checks worth running before you move real traffic. Each takes a couple of minutes and none require payment.

Is there a price list before registration

If prices appear only after you create an account, that is an answer in itself. The normal practice is a publicly available model list and prices. At Zukko /v1/models returns the identifiers, while prices and context are at /v1/models/info (and on the Models page in the dashboard).

curl https://api.zukko.pro/v1/models
curl https://api.zukko.pro/v1/models/info

While you are there, look at what the endpoints actually return: they should carry the identifiers you will put in your config and prices in comprehensible units.

Is the rate symmetric

With most providers, output tokens cost three to five times more than input. Intermediaries sometimes differ, and that changes the arithmetic.

The rule is simple: for agent workloads look at the input price, for chat workloads at the output. A flat rate favours agents; an asymmetric one favours short prompts with long answers.

Does billing match usage

The most important check, and it happens on your very first request.

The API returns the exact token count in the usage field:

"usage": {"prompt_tokens": 6499, "completion_tokens": 214}

Make a request, note the numbers, open the usage history. Different figures there mean you can stop looking. If they match, also check whether failed requests are billed: a provider's error should not cost you money.

Is the protocol compatible

If the API matches the original, moving back costs a two-line edit. That is your insurance against the intermediary raising prices or disappearing.

Test it like this: take your existing code, change the base URL and key, run it. If anything else needed changing, compatibility is partial and the dependence is deeper than it looks.

What happens when the provider fails

Ask directly, or test it: what comes back when a model is unavailable. A good answer is a clear error with a code your code can act on. A bad one is someone else's message with someone else's support contacts — or worse, a 200 with an empty body.

Check streaming separately: at some gateways a streamed error arrives with a 200 status inside an event, and a naive client treats that as success.

Can you get your money back

A balance you cannot withdraw is not a balance, it is a prepayment with no obligations attached. Clarify the terms before putting in more than a trial amount.

And keep only what you can afford to lose on the account: a sensible habit with any intermediary is topping up as you spend rather than depositing a yearly budget at once.

A short checklist

Migrate in two steps: staging with a full real task first, production after. A short test request will not surface problems with long context, tools or streaming.