ChatGPT API in Python

ChatGPT’s paid API can be used in Python to build custom, stateful chat applications—by sending a growing list of prior “messages” (user and assistant turns) with every request. The core workflow is straightforward: create an OpenAI account, set up billing, install the updated OpenAI Python package, load an API key, then call a chat-completions endpoint using a chosen model such as gpt-3.5-turbo. Responses return structured fields, and the practical output is the assistant’s text inside completion.choices[0].message.content. The cost is low (described as about $0.002 per thousand tokens), but token limits matter because requests stop once they hit a maximum length (noted as 4096 tokens in the example).

A key operational detail is that the API does not automatically remember conversation history. To get multi-turn behavior, developers must maintain message_history themselves: append each user input and each assistant reply, then resend the entire history list on the next call. The tutorial demonstrates this by first asking a factual question about the Moon’s circumference, then asking a follow-up that depends on context (“which moon is that in reference to?”). The follow-up produces inconsistent behavior—sometimes failing to use prior turns as expected—highlighting that model behavior can be finicky and may change across runs or model versions. The takeaway is less about “bugs” and more about the reality that prompt formatting, history handling, and model updates can affect reliability.

To make the pattern reusable, the tutorial refactors the repeated request logic into a chat function that takes user input, appends it to message_history, calls the API, extracts reply content, appends the assistant message back into history, and returns the reply. This same structure supports practical safety-oriented Q&A, such as asking whether water collected by a dehumidifier is safe to drink and then requesting ways to make it safer. The assistant provides guidance (e.g., boiling and filtration), while also noting limitations like boiling not removing certain chemicals that could leach from the device.

For a user-facing interface, the tutorial then integrates the API with Gradio to build a simple chat UI. It constructs a “joke bot” by adding a pre-prompt message that instructs the assistant to respond only with jokes related to the user’s subject. The Gradio app uses a predict function wired to a text box submit event, updates the conversation history, and returns the chat transcript in the format Gradio expects. In testing, the jokes vary in quality, but the interface demonstrates how easily the same API can power either constrained behavior (like joke-only responses) or general Q&A.

Overall, the central insight is that building with the ChatGPT API is mostly about managing messages, choosing models, and handling token limits—then layering on whatever UI or behavior constraints the application needs. Even with strong capabilities, factual accuracy can still fail, so outputs should be treated as fallible guidance rather than ground truth.