GPT API relay services 101 in total

GPT has the widest coverage of any family — more relays support it than anything else, because OpenAI's API specification has effectively become the industry standard and is the cheapest to integrate.

The distinction most often overlooked is between "Chat Completions compatible" and "Responses API supported". The former is the common standard and nearly every service implements it; the latter is the newer shape that tools such as Codex CLI depend on. A service with only the former fails outright with the new tools — which explains the familiar puzzle of a configuration that looks entirely correct and still does not work.

Another practical issue is how current the model list is. The GPT line iterates quickly, so a /v1/models response that only lists model names from years ago usually means the service has gone stale. Our probe records the actual response, so you can compare directly.

Two billing points to keep in mind: a reasoning model's thinking also counts as output tokens, so the bill runs well above a per-conversation estimate; and cache support varies a great deal between services, which can make long-context costs differ several times over.

• Most stable：Highest current uptime. Where samples are thin we show the public-dataset baseline instead, and the detail page states the source.
• Fastest：Home-page response time from the most recent probe. A single measurement is noisy, so treat it as a range rather than an absolute.
• Broadest model coverage：This uses the model count our endpoint probe actually detected, not the support list the service advertises — the two are measured differently, and the advertised figure is usually the more optimistic one.

After switching base_url to a relay, does my code need changes?

In most cases you only change base_url and api_key; SDK code stays as it is — that is the value of OpenAI-protocol compatibility. But if you use newer endpoints (Responses API, Realtime), confirm the service implements them first, or you will get "basic chat works, this feature 404s".

Are GPT models downgraded on relays?

It happens, though less often than with the Claude line. The checks are similar: look at what /v1/models returns, compare answer quality, and see whether latency matches the tier.

Why does the same model run at such different speeds on different services?

Two things dominate: the network path from the service to its upstream, and whether it queues requests. We collect latency daily, so the distribution is visible on the comparison board. Single measurements fluctuate a lot — read a period's median rather than one reading.

Are reasoning models a good fit for relays?

They can be, but watch the billing. A reasoning model's thinking also counts as output tokens and often far exceeds the final answer, so the bill runs well above a per-conversation estimate. Run a week of real tasks after signing up, then re-evaluate from actual token usage.