Gemini API relay services 63 in total

Gemini's differentiating strength is long context and multimodality. A million-token context window makes it practical to drop in an entire codebase or analyse a long video directly — capabilities the other two families cannot easily replace yet.

The critical question for these services is whether the long context is real. Printing a window size in the docs is easy; actually handling a million tokens puts demands on the upstream channel and on billing. The test is straightforward — feed it a long document close to the limit, and see whether it is handled completely and whether it is quietly truncated past some length.

Multimodality works the same way. A page may claim image, video and audio understanding, but only real material proves it — video input in particular asks more of the channel, and some services fail silently with a "not supported" response.

How billing is defined is worth attention too. Long-context requests carry a lot of tokens, so if a service re-forwards them and breaks caching, re-sending the same conversation pushes the cost up sharply. For long-document work, prefer services that support context caching.

• 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.

Can a relay really handle a million-token context?

Not always. Some services list Gemini Pro but cap input at a much lower length, returning an error or truncating silently beyond it. Test with a document close to the stated limit and plant a question inside it that can only be answered by reading to the end.

Flash or Pro — which should I pick?

Split by task complexity. Flash is cheap and fast, a good fit for classification, extraction and format conversion; Pro suits work that needs reasoning and long-context comprehension. In practice, run everything on Flash first and move only the traffic that falls short onto Pro — that is the cheapest arrangement.

Are video and audio inputs reliable through a relay?

Less predictable than image input. Image support is common; video and audio are implemented by only some services. If you depend on them, run a full test with real material first rather than building against a feature list.

How do I keep costs down in long-context work?

The key is making context caching take effect: put the fixed long document at the front of the request and the changing parts at the back. That matches the vendor's own guidance, and it depends on the relay passing the cache fields through. Our detail pages record each service's probe results, which is a reasonable place to start filtering.