# LLM API Tester > Free in-browser tester for OpenAI-compatible LLM APIs, relays and proxies. Sends one prompt to up to three chat completions endpoints in parallel and reports the reply, HTTP status, the model name the server returns, time to first token, total time, output tokens per second and token usage. Independent tool, not affiliated with any API provider. Updated 2026-11-18. Full page: https://lab.awaara.cn/llm-api-tester/ ## What it does - Inputs per endpoint (up to 3): base URL (normalized: https:// added when missing, /v1 added when the path is empty, a trailing /chat/completions or /models removed, each change reported), API key (password field), model name, and a List models button that calls GET {base}/models and fills a picker. - Shared inputs: prompt, max output tokens (1 to 8,192, default 512; out-of-range values are clamped and the page says so), stream on or off. - Request: POST {base}/chat/completions with stream: true and stream_options: {"include_usage": true}. If the server rejects stream_options it retries without them; if it rejects streaming it retries without streaming and says so. 60-second limit per endpoint, with a Stop button. - Results per endpoint: reply text, HTTP status, returned model (marked as a match, a dated version of the requested name, or different), time to first token, first answer token for reasoning models, total time, output tokens per second, prompt and completion tokens (n/a when the server sends none; speed is then estimated at about 4 characters per token and marked as an estimate). - Comparison: with 2 or 3 endpoints, a side-by-side table of TTFT, total time, tokens per second and returned model. - Errors: CORS or network failure (explained, with a ready curl command that uses $API_KEY instead of the real key), 401/403, 404 (model or path), 429 with Retry-After when readable, 4xx and 5xx with the server's message, timeouts, and dropped streams. ## Definitions - Time to first token (TTFT): from sending the request to the first streamed token, including network round trip, queueing and prompt processing. Needs streaming. - Output tokens per second: completion tokens divided by the time between the first and last streamed token. Without streaming: completion tokens divided by total time. - Returned model: the model field in the response. It is a label written by the server and does not prove which model answered. Use the Model Fingerprint Lab (https://lab.awaara.cn/model-fingerprint/) to test the model itself. ## Privacy - Requests go directly from the browser to the endpoint the user enters, with fetch. No proxy; lab.awaara.cn never receives the key, prompt or reply. - Keys are kept in memory only, unless the user ticks "Remember on this device", which saves base URLs, keys and model names in the browser's localStorage until unticked. - Analytics (Umami) records only the outcome category (ok, cors, http-4xx, http-5xx, timeout), the number of endpoints and whether streaming was used. Never URLs, keys, model names or prompts. ## Limits - OpenAI-compatible chat completions only. Gemini works through https://generativelanguage.googleapis.com/v1beta/openai and Claude through https://api.anthropic.com/v1 (OpenAI SDK compatibility), if the provider allows browser (CORS) requests. - Relays that send no CORS headers cannot be tested from any web page; use the generated curl command instead. - Timings are one request from the user's connection at one moment; run several times before comparing. ## Sources - OpenAI, Create chat completion (stream_options.include_usage): https://platform.openai.com/docs/api-reference/chat/create - Google AI for Developers, OpenAI compatibility: https://ai.google.dev/gemini-api/docs/openai - Anthropic, OpenAI SDK compatibility: https://docs.anthropic.com/en/api/openai-sdk - MDN, Cross-Origin Resource Sharing (CORS): https://developer.mozilla.org/en-US/docs/Web/HTTP/Guides/CORS