OpenAI retires its last GPT-3.5 models today

Four model IDs stop answering today. gpt-3.5-turbo-instruct, babbage-002, davinci-002 and gpt-3.5-turbo-1106 reach shutdown on September 28, 2026 per OpenAI's deprecations page, all pointed at gpt-5.6-terra as the replacement. If you are reading this after a 404 in your logs, that is the story. But the OpenAI model deprecation worth clearing your calendar for is October 23, which is four weeks out and takes fine-tuned models with it.

What actually happened

Today's four were announced for deprecation on September 26, 2025 — 367 days of notice. That is the generous end of OpenAI's range, and it is not the norm. gpt-5.4-cyber shuts down October 1 on an announcement dated September 11, 2026: twenty days.

The October 23 batch is the one that matters. It retires gpt-3.5-turbo-0125 and its aliases, gpt-4-0613, the gpt-4-turbo variants, o1 and o1-pro, and gpt-image-1 (replaced by gpt-image-2). Recommended replacements are gpt-5.6-terra, gpt-5.6-sol or gpt-5.6-luna depending on the original tier.

It also retires fine-tuned models — ft-gpt-3.5-turbo and ft-gpt-4. That is a different kind of problem from swapping a string. A fine-tune is a base model plus your training run. When the base disappears, the artifact you paid to produce disappears with it, and you do not re-point it — you re-run it against a new base and re-validate the output.

OpenAI's doc says deprecated models "will no longer be accessible," with a note that some organizations may explore provisioning dedicated capacity for continued access. Read that as: there may be an enterprise escape hatch, and it is not a line item you want to discover on October 23. All of this lands the day before DevDay on September 29, which will almost certainly add new models to the same list.

Why model deprecation matters for your business

The notice window is the vendor's choice, not a standard. 367 days for one batch and 20 for another, from the same vendor, in the same month. Any plan that assumes "they always give us a year" is assuming a courtesy, not a contract. If you need a guaranteed window, it belongs in a commercial agreement or you accept the risk explicitly.

Fine-tunes are perishable assets. Teams treat a fine-tuned model like a build artifact — produced once, deployed forever. It is closer to a leasehold. Keep your training data, your eval set and your run configuration in version control alongside the code, so re-creating the model on a new base is a pipeline you execute rather than an archaeology project.

Grep is your migration audit. Model IDs get hardcoded in the places nobody looks: a cron job, a one-off script, a Zapier step, an old Lambda, a notebook somebody scheduled. Search every repo, every environment variable store and every no-code tool for gpt-3.5, gpt-4-0613, gpt-4-turbo, o1 and gpt-image-1 before the 23rd, not after.

A newer model is not a drop-in. gpt-5.6-terra is not gpt-3.5-turbo-instruct with better scores — different tokenization behavior, different instruction-following, different refusal boundaries, different cost per call. Prompts tuned against a 2023 model routinely need rework. Run your eval set against the replacement before you flip the switch, and budget for prompt edits.

One config line is the whole defense. Every system we build puts the model ID in configuration and every call behind a single interface. When a deprecation notice lands, the change is one line and a test run. When it is scattered across forty files, it is a sprint.

Sources: OpenAI model deprecations, Announcing OpenAI DevDay 2026.