Should you use ChatGPT or text-davinci-003 in the Playground?

Two new things from OpenAI arrived within days of each other this month. On November 28, the company added text-davinci-003 to its API, the paid service developers use to build language models into their own software. Two days later it opened ChatGPT, a free chatbot, to everyone. Both come from the same broad family of models, so people who have tried both often ask which one to use.

The short answer depends less on which model is "smarter" and more on how you want to work. ChatGPT is a conversation. The Playground is a workbench.

What each one is

ChatGPT runs at chat.openai.com. You type, it replies, and it keeps the earlier turns of the conversation in view, so you can say "shorter" or "now in French" without repeating yourself. It is free while OpenAI calls it a research preview, and it offers almost no settings.

text-davinci-003 is a model you reach through the API or through the Playground, a web page in your OpenAI account where you type a prompt into a large text box and the model continues it. OpenAI's release note says it builds on its earlier InstructGPT models (versions of GPT-3 trained to follow instructions) and improves on text-davinci-002 in three ways: higher-quality writing, better handling of complex instructions and better long-form text. OpenAI's own prompt guide names it the best option for text generation as of November.

The Playground is not free in the same way. Davinci models cost $0.02 per 1,000 tokens, counting both your prompt and the reply, and a token is roughly four characters of English text, so 1,000 tokens is about 750 words. New accounts get $18 in free credit for the first three months, and OpenAI's pricing page states that Playground use counts against your account just like any other API call.

A worked same-prompt test

The fairest way to choose is to give both the same job. Here is a simple procedure.

1. Pick a task you actually need, for example:

Write a 120-word product description for a handmade maple cutting board. Audience: home cooks who care about where things are made. Mention care instructions in one sentence. Plain, warm tone.

1. Paste it into ChatGPT and save the reply.
2. In the Playground, choose text-davinci-003, set the temperature to 0.7 and the maximum length to 250 tokens, paste the same prompt and press Submit.
3. Run each one twice more. In ChatGPT, ask for a new response; in the Playground, submit again.
4. Score all six versions on the same three questions: Did it follow every instruction? Is the length right? Would you publish it with light edits?

Here is the kind of comparison you might end up with. It is illustrative; your results will vary.

For one-off writing and learning, many people will prefer the chat. For a task you plan to repeat a hundred times, such as product descriptions for a whole catalog, the Playground's settings and the path to the API matter more.

What can the Playground do that the chat can't?

The Playground gives you the controls behind the model's answers. OpenAI's prompt guide says the model and the temperature are the settings people change most. Temperature controls how often the model chooses a less likely word: higher values give more varied, often more creative text, while 0 gives the most predictable output, which OpenAI recommends for factual tasks such as pulling data out of a document. Varied is not the same as accurate, so a high temperature does not make answers more truthful.

Maximum length sets a hard limit on how many tokens the reply may use. It is a ceiling, not a target: the model stops earlier if it decides it is done, and if the cap is too low the text is cut off mid-sentence. Stop sequences let you name a piece of text, such as a line break or "###", that ends the reply as soon as it appears, which is useful when you want exactly one answer in a fixed format. The Playground also lets you pick between models, from the most capable Davinci to faster, cheaper ones, and shows a running token count so you can see what each prompt costs.

The other difference is transparency. In the Playground, the model sees only what is in the text box, so you know exactly what it was given. That makes results easier to repeat and easier to turn into code later. ChatGPT, in exchange for fewer controls, remembers the conversation and makes revisions effortless. Many people will use both: the chat to explore an idea, the Playground to make a version they can run the same way every time.