ModelRefs / Text Completion — AI Glossary

Text Completion — AI Glossary

A legacy LLM API endpoint accepting a raw text prompt and returning a continuation; superseded by the chat completion format.

Overview

Text completion (POST /v1/completions) was the original GPT-3 API interface. It accepts a plain text prompt and returns a completion string. No role structure—suitable for base model prompting and fine-tuned classification models. Deprecated by OpenAI for most use cases in favor of chat completions. Still used for legacy integrations and raw base model access.

Reference details

Topic	ecosystem
Last reviewed	2026-06-24

How it works

The request carries a single string and the response carries its continuation, with no message array and no role structure anywhere in the exchange.
That shape matches what a base language model actually does, which is extend a token sequence, so nothing in the interface implies a conversation.
Because there are no role delimiters, the prompt is exactly the tokens the model receives, which makes few-shot formatting fully explicit and fully the caller's responsibility.
Stop sequences do the work that a turn boundary does in a chat interface, terminating generation when a caller-specified string appears.
Instruction following is not assumed: a base model served this way continues the pattern it was given rather than responding to a request.
Chat interfaces are built on top of this behaviour, serialising a message array into exactly such a string using the delimiter tokens the model was tuned on.

Example: Where the framing overhead is worth avoiding, and where it vanishes

Chat serialisation adds role delimiters around each message, on the order of 4 tokens per message, so a three-message request carries roughly 12 tokens of framing. What that costs depends entirely on the prompt it is added to. On a 50-token classification prompt it is 12 / 50 = 24% overhead; on a 2,000-token prompt it is 12 / 2,000 = 0.6%, which is to say nothing. Priced out on the short case: a 45-token prompt returning a single token costs 0.000045 x $3 + 0.000001 x $15 = $0.00015 raw, and 57 input tokens after framing give 0.000057 x $3 + 0.000001 x $15 = $0.000186, a ratio of about 1.24. Across 50,000,000 classifications that is $7,500 against $9,300, a difference of $1,800. Whether $1,800 justifies maintaining a deprecated code path is a real question with a defensible answer either way, and it is not the main reason the endpoint persists. The structural reason is access: a chat interface wraps the prompt in delimiters the model was instruction-tuned on, so it cannot present a base model with a raw sequence. For base-model research, for log-probability work over exact token strings, and for fine-tuned classifiers that were trained on unwrapped prompts, the completion interface is not a cheaper option but the only one that sends what was intended.

Commonly confused with

Text completion is not a worse chat completion, and treating it as a deprecated predecessor misses what it is for. The two interfaces expose different things: chat exposes an instruction-tuned model through the role structure it was trained on, while completion exposes raw continuation. For an instruction-tuned model the chat interface is strictly better, since matching the training format is what produces good behaviour. For a base model there is no equivalent, because wrapping a raw sequence in chat delimiters changes the input. The second confusion is with the completion of a chat turn, sometimes called prefilling, where an assistant message is begun and the model continues it. That is a chat-interface feature using the role structure, not a use of this endpoint.

When to use it

Reach for it when:

For base-model access, where chat delimiters would alter the sequence being sent.
On fine-tuned models trained on unwrapped prompts, which expect exactly the format they were tuned on.
For log-probability work over an exact token string, where any added framing changes the measurement.
On very short high-volume prompts, where framing was 24% overhead against 0.6% on a long one.

Reach for something else when:

With instruction-tuned models, which behave best through the role structure they were trained on.
For anything multi-turn, where the caller would be reimplementing conversation state the chat interface maintains.
For tool use or structured output, which are expressed through the message and role structure this interface lacks.
As a cost optimisation on long prompts, where the framing overhead was 0.6% and not worth a deprecated code path.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Text Completion — AI Glossary.

Frequently asked questions

What is Text Completion?