From Completions to Conversation

The Mystery of chat/completions

We're clearly having a conversation with a large language model — so why does OpenAI's API path say "completions"? The answer to this naming choice reveals the true nature of LLMs.

An Interesting Question

It's clearly a "chat" — so why is it called "completions"?

Two companies, two very different naming choices — each reflecting a different understanding of what LLMs fundamentally are.

OpenAI
/chat/completions
Chat / Completions
Anthropic
/v1/messages
Straightforward messages
OpenAI's Implicit Logic
The essence of an LLM is completion: give it some text, and it continues writing. Chat is simply dialogue formatted as "text to be completed" — it's still Token completion at its core.
The Key Point for PMs
Every message you send to an LLM gets assembled into a single block of text, and the model then continues generating from the end of that text. Conversation is just a user-friendly shell humans have added on top.
Viewing a Conversation Through the Completions Lens
What the model actually "sees" (the assembled text)
[system] You are a helpful assistant.
 [user] Who is Zixia Fairy? (紫霞仙子)
 [assistant]
What the Model Does
After the [assistant] marker, it predicts the next highest-probability Token, continuing until it generates a stop token
Chat Is an Illusion
You think you're "having a conversation" — the model is simply "completing this block of text." That's the truth behind "completions"
Why PMs Need to Know This
System Prompt, context, and message history are all concatenated and sent to the model together — Token count is the key lever for both cost and quality
OpenAI's Design Decision
This "literal" naming choice
shaped the entire world.
GPT + /chat/completions was the paradigm that first ignited the market
Developers, frameworks, and toolchains worldwide all aligned to this interface design
Understanding "conversation = completion" is the starting point for making sense of all AI engineering