What the API is used for

Conversation, instruction-following and function calling — the workload most people mean by 'an LLM API'. This is the drop-in case: point an existing OpenAI client here and change the model id.

Models serving it: GLM-4.5-Air ($1.50 / 1M tokens), Qwen-2.5 ($0.15 / 1M tokens).

What it is not. Not for tasks that need a frontier closed model's reasoning. Where one is genuinely better, the cheaper request is not the better one.

Generate video from a text prompt or an input image.

Models serving it: Wan2.1-VACE Video ($50.00 / 1M tokens).

What it is not. Slow and expensive relative to everything else here. Worth checking the rate on the pricing page before designing around it.

Send an image with your prompt and get text back — reading a screenshot, pulling fields off an invoice or a form, describing a chart. Same chat-completions shape, with an image part in the message.

Models serving it: Qwen-3.5 ($0.15 / 1M tokens).

What it is not. Not an OCR engine. It reads a document to answer a question about it; it does not produce a faithful character-level transcription of a scan.

Turn recordings into text — meetings, interviews, voice notes, call audio. Billed per minute of audio rather than per token, so the cost of a job is knowable before you run it.

Models serving it: Whisper Large v3 ($0.004 / minute).

What it is not. Not real-time streaming transcription. You upload a file and get a transcript; there is no live socket.

By what you are sending, not by price — a speech model cannot answer a chat request, so most of the choice is made for you. Within a workload, run the cheaper model first and move only if its output is not good enough for yours. That order costs you one experiment; the reverse costs you the difference on every request in the meantime.

Every model, rate and context window is on the catalogue and the pricing page. What each endpoint accepts is in the API reference, and what it returns when something is wrong is on the errors page.