Run a test
Each run sends one chat completions request per endpoint, straight from this page, with a 60-second limit. Streaming is needed to measure time to first token. If a server rejects streaming, the tester retries once without it and says so.
What it measures
- Time to first token (TTFT): from the moment the request leaves your browser to the first streamed token. It includes the network trip, queueing and prompt processing.
- Total time: until the last byte of the reply arrives.
- Output tokens per second: completion tokens divided by the time between the first and the last token, so the wait before the first token does not drag it down. Without streaming, it is completion tokens divided by total time.
- Returned model: the
modelfield in the response, compared with the name you asked for. - Usage: prompt and completion tokens as the server reports them. When the server sends none, the page says n/a and estimates speed from characters instead.
The numbers describe one request from your connection at one moment. Run it two or three times before you judge a relay: the first request often pays for a fresh TLS connection, and providers have busy minutes.
Why the returned model can differ
Official APIs often answer with a dated snapshot of the alias you asked for, such as a name with a release date on the end. The tester treats that as a match and says so. A completely different name is another story: some relays route requests to a cheaper model to save money, and a few forget to rewrite the field when they do.
The field is only a label the server writes, so a matching name proves nothing on its own. To check the model itself, run the Model Fingerprint Lab against the relay and the official API and compare the results. Which model is behind an API? covers six checks, from asking the model to token counts.
Your key stays in your browser
The page sends each request with fetch from your browser directly to the base URL you typed. There is no proxy, so this site never sees your key, prompt or reply. You can confirm it in your browser's network panel: the only request a run makes goes to your endpoint.
Keys are kept in memory and gone when you close the tab. If you tick Remember on this device, base URLs, keys and model names are saved in this browser's local storage until you untick the box. Anyone who can use this browser profile can read them, so leave it off on shared computers. A key with a low spending limit is the safest kind to test with.
The page counts runs with Umami, which records only the outcome (such as ok or timeout), the number of endpoints and whether streaming was used. It never records URLs, keys, model names or prompts.
FAQ
Why does my relay fail here but work in curl?
A browser only lets a page read a response from another site when that site sends CORS headers that allow it. curl and server code ignore CORS, so a relay that never sends those headers works from a terminal and fails in every browser. When that happens, the tester shows a ready curl command with $API_KEY in place of your key, so you can run the same request yourself.
Is TTFT the same as latency?
No. Time to first token is the wait from sending the request to the first streamed token. It includes the network round trip, queueing and prompt processing. Total time runs to the last token, so it also grows with the length of the reply. A reasoning model that streams its thinking gets a first token early; the tester then also shows when the first answer token arrived.
Does this work with Anthropic or Gemini?
Only through their OpenAI-compatible endpoints, because the tester speaks the chat completions format. For Gemini use https://generativelanguage.googleapis.com/v1beta/openai as the base URL; for Claude use https://api.anthropic.com/v1. Whether a browser may call them directly depends on each provider's CORS policy; if it fails here, the curl command still works.
Why does usage show n/a?
The server did not send a usage object. When streaming, the tester asks for one with stream_options: {"include_usage": true}, which makes compliant servers add a final chunk with token counts. Some relays strip it or do not support it. Tokens per second is then estimated from the characters received, at about four characters per token, and marked as an estimate.
Does a matching model name prove which model answered?
No. The model field is a label the server writes, and a relay can write any name it likes. A mismatch is worth a closer look, but a match proves nothing. To test the model itself, compare its statistical fingerprint against the official API with the Model Fingerprint Lab.
Sources
OpenAI, Create chat completion (API reference): the request format, streaming, and stream_options.include_usage adding a final usage chunk before [DONE].
Google AI for Developers, OpenAI compatibility: the Gemini base URL for OpenAI-format requests.
Anthropic, OpenAI SDK compatibility: the Claude base URL and support for stream and stream_options.
MDN, Cross-Origin Resource Sharing (CORS): why a browser blocks responses that curl can read.
Agents
Machine readable summary for AI assistants: