GPT-3.5 Turbo Guide: The 23 October Shutdown and How to Move

If anything in your stack still sends model: "gpt-3.5-turbo", the most useful fact in this article is a date. OpenAI's deprecations page lists 23 October 2026 as the shutdown for gpt-3.5-turbo-0125, and with it the aliases gpt-3.5-turbo and gpt-3.5-turbo-completions. As of this writing that is six days away.

The deprecation was announced on 22 April 2026, so this is not a surprise to anyone reading OpenAI's changelog. It is a surprise to the cron job nobody has touched since 2023, and that is usually where these calls live.

What GPT-3.5 Turbo is, and why it lasted

GPT-3.5 Turbo was OpenAI's chat-optimised GPT-3.5 model, introduced at a tenth of the cost of text-davinci-003, and for a long stretch it was the default cheap model for anything conversational. OpenAI's model page now calls it a legacy model for cheaper chat and non-chat tasks. The current snapshot has a 16,385-token context window, a 4,096-token output cap and a knowledge cutoff of September 2021.

Its pricing on OpenAI's page is $0.50 per million input tokens and $1.50 per million output. That price, and the fact that it was good enough for classification, short summaries and simple chat, is why it stayed in so many codebases long after better options arrived.

OpenAI has recommended moving off it since July 2024, when its model page started pointing to gpt-4o-mini as cheaper, more capable, multimodal and just as fast. Two years of that advice did not move every caller. A shutdown date will.

Snapshots, aliases and what each name means

The family has accumulated several names, and they have different fates:

• gpt-3.5-turbo — an alias that points at the current snapshot, gpt-3.5-turbo-0125. Retiring on 23 October 2026.
• gpt-3.5-turbo-1106 — an earlier dated snapshot, retired on 28 September 2026 alongside the instruct model.
• gpt-3.5-turbo-0613 — the June 2023 snapshot. OpenAI shut it down on 13 September 2024. The suffix is the snapshot date; dated names exist so you can pin behaviour while the alias moves.
• gpt-3.5-turbo:batch — an aggregator endpoint listed on OpenRouter at half the normal price, $0.25 in and $0.75 out per million tokens at the time of writing. Same model, run asynchronously, which only pays when nobody is waiting on the answer.

The two siblings have their own guides: GPT-3.5 Turbo 16k and GPT-3.5 Turbo Instruct, the latter already switched off at OpenAI.

One caution about aggregators. A catalogue may keep listing a model after the vendor retires it, and whether a request still succeeds depends on what sits behind the listing. Do not treat a catalogue entry as evidence the model is still served.

Find every caller before you change one

The hard part of a migration like this is not choosing the replacement. It is finding the calls. A practical search order:

1. Grep source, config and infrastructure repositories for gpt-3.5, including environment files and Helm values.
2. Check your provider's usage dashboard for requests by model over the last 30 days. Anything that shows up there and not in your grep is running from somewhere you forgot.
3. Look at no-code tools, spreadsheets with scripts attached, and third-party integrations configured through a web UI. These rarely live in git.
4. Check fine-tuned models. A fine-tune built on a retiring base is a separate migration, because the training has to be redone on the new base.

Picking the replacement

OpenAI's deprecations page names gpt-5.6-terra as the recommended replacement; the GPT-5.6 Terra guide covers that model. Whatever you pick, do not assume the cheapest current model behaves like GPT-3.5 Turbo. Newer models follow instructions more literally, write longer answers by default and format differently, and any downstream parser written against 2023 output can break on all three.

The safe procedure is short:

1. Pull 50 to 100 real requests from logs, with the responses GPT-3.5 Turbo gave.
2. Replay them against the candidate with the same prompts and parameters.
3. Diff for the things your code depends on: JSON validity, length, refusals, label sets in classification tasks.
4. Fix prompts where needed, then switch the model name and pin the new dated snapshot.

If the task was classification or extraction, consider using structured outputs on the new model rather than reproducing the old free-text parsing. It removes a whole class of failure the old setup had to tolerate.

What not to do this week

Do not point the old name at a fallback in your router and call it done; a silent fallback hides the fact that behaviour changed. Do not migrate everything to the most expensive model out of caution either. Most GPT-3.5 Turbo workloads were simple, and a small current model is usually enough — measure it.

And once you have moved, pin the dated snapshot of the new model rather than the bare alias. Planning for deprecation covers how to make the next one of these a scheduled task instead of an incident.