OpenAI API

At a glance

The OpenAI API gives developers hosted models for text, code, images, audio and tool-using applications. You send requests from your own software and receive model output; you still build the interface, data flow and checks around it. It is useful when you want model capabilities without operating model-serving hardware.

Good fit

A practical candidate for adding generation or interpretation to an existing product. Start with one narrow task and judge representative outputs against your own requirements. Model choice, prompt length, response length and tool use affect both cost and behaviour, so a good fit is something to test rather than infer from a model name.

Consider something else when

Your requirement is fully offline execution or keeping all inputs off a hosted provider. You also need your own validation before treating generated answers or tool arguments as authoritative. A hosted endpoint removes model-serving work; it does not establish correctness for your application.

Start here

Create a project key and keep it in a server-side OPENAI_API_KEY environment variable. With the official SDK, make one small Responses request through client.responses.create using a model available to your organization. Inspect output_text and token usage, then evaluate a few representative cases before building an automated action around the result.

Cost and requirements

Usage pricing varies by model, tokens and tools. As an October 30 example, standard short-context gpt-6.1-sol lists $2 per million input tokens and $10 per million output tokens; cache, longer-context and tool charges can differ. Eligible paid ChatGPT Pro or Business owners can claim linked API credits, after seven days, subject to ownership, activation and expiry rules. The benefit excludes Codex API-key usage.

Limits to know

API inputs and outputs are not used for training by default unless you opt in, but that does not mean zero retention. Default abuse-monitoring logs can be kept for up to thirty days, and Responses can retain application state. Zero Data Retention requires approval and has feature restrictions. Review those data boundaries alongside organization limits before sending sensitive production inputs.

Related work

• capy-flash: local inference for selected Qwen models on supported hardware (alternative)