why does the openai api answer differently than chatgpt?

Also the ChatGPT app renders markdown, so headings and bullet lists look neat there and come back as raw asterisks and hashes over the API. Sometimes people think the answer is worse when it's just not formatted.

ChatGPT isn't the bare model. It wraps it in its own system prompt, may use memory from your past chats, and can switch on tools like web search, none of which your API call gets unless you add it. The model name and default settings can differ too. On the API you're starting from nothing, so write a system message describing how you want it to answer and lower the temperature if you want more consistent output. It still won't match ChatGPT word for word, but it gets close enough.