SunsettingAI Dev Tools

gpt-3.5-turbo-instruct, babbage-002, davinci-002, gpt-3.5-turbo-1106

by OpenAI

Announced

2025-09-26

Complete

2026-09-28

Refund window closes

—

Refund status

No refund applicable

Four legacy OpenAI model snapshots lose API access on September 28, 2026: `gpt-3.5-turbo-instruct`, `babbage-002`, `davinci-002` and `gpt-3.5-turbo-1106`. OpenAI announced the batch on September 26, 2025 under the heading "Legacy GPT model snapshots" and lists `gpt-5.6-terra` as the recommended replacement for all four. Three of the four are themselves the models OpenAI handed out as replacements in the January 4, 2024 retirement of the original GPT-3 and InstructGPT lineup, so anyone who migrated then and pinned the new name is being migrated again now. Requests naming these model strings stop working on the shutdown date rather than degrading, so the failure shows up as an API error, not as worse output.

Refund flow

1
Not applicable. These are pay-per-token API models with no pre-paid credit balance attached to a specific model string, so there is no refund category to open.
2
If you are billed for one of these four model strings for usage dated after September 28, 2026, treat it as a billing error rather than a refund request: open a ticket at help.openai.com with the invoice line and the request IDs.
3
The cost worth checking before the date is not a refund, it is a per-token price difference. Compare your current spend on these models against the replacement on OpenAI's pricing page before you switch, because the older base models were priced for a different generation of hardware.

Migration path

Recommended successor

gpt-5.6-terra

Why this is the closest fit

It is the replacement OpenAI names itself, on the same row as each of the four retiring models, on its own deprecations page.

What differs from the original

The bigger break is the endpoint, not the model. `gpt-3.5-turbo-instruct`, `babbage-002` and `davinci-002` are completion-style models, called with a single `prompt` string and answered with a raw continuation. The replacement is a chat and responses-era model, called with structured input and system instructions. Code that sends one flat prompt and reads `choices[0].text` needs both a different request shape and a different response path. Behaviour differs too: a completion model continues your text, while an instruction-tuned chat model answers it, so few-shot prompts written as bare text continuations often need rewriting as explicit instructions. `gpt-3.5-turbo-1106` is the easier case, since it was already a chat model and the change is mostly the model string.

Alternatives by use case

You called gpt-3.5-turbo-instruct with a flat prompt string

gpt-5.6-terra via the Responses API, rewriting the prompt as an instruction rather than a text continuation

You pinned gpt-3.5-turbo-1106 for output reproducibility

gpt-5.6-terra, and pin the dated snapshot rather than the alias so the next alias update does not move your outputs

You used babbage-002 or davinci-002 as a cheap classifier or scorer

Benchmark gpt-5.6-terra on your own labelled set before switching, because base-model log-probability tricks do not transfer cleanly to an instruction-tuned model

You fine-tuned babbage-002 or davinci-002

Those fine-tuned variants have a later date: OpenAI lists ft-babbage-002 and ft-davinci-002 as shutting down October 23, 2026, so inference on your existing fine-tune outlives the base model by about four weeks

What this tool meant

This row is the end of the GPT-3 completion era at OpenAI. The models retiring here are not the originals, they are the survivors: on January 4, 2024 OpenAI shut down `text-davinci-003`, `text-davinci-002`, `text-curie-001`, `text-babbage-001`, `text-ada-001` and the bare `ada`, `babbage`, `curie` and `davinci` base models, and pointed users at exactly these replacements. `gpt-3.5-turbo-instruct` absorbed the InstructGPT traffic; `babbage-002` and `davinci-002` absorbed the base-model traffic. Twenty months later the replacements are being retired in turn, which is the pattern worth noticing: a migration target is only as durable as the provider's interest in the endpoint shape behind it, and the completions endpoint has been the least-loved shape at OpenAI since chat arrived. The practical lesson for anyone running production calls against a hosted model: the deprecation page is the only contract, notice periods are counted from an announcement you have to be watching for, and a model named as a replacement is not a promise of permanence. OpenAI gave twelve months here, which is generous by the standards of its own policy minimum of six months for generally available models, and it still costs a rewrite rather than a string swap because the endpoint changed underneath.