GPT-3.5 Turbo Retirement —
deprecation dates & what replaces it
Quick answer: GPT-3.5 Turbo is a legacy model family. The dated snapshot gpt-3.5-turbo-0613 was retired on September 13, 2024, and gpt-3.5-turbo-0301 was deprecated on June 13, 2023. OpenAI's recommended replacement for GPT-3.5 Turbo workloads is GPT-4o mini — cheaper per token than GPT-3.5 Turbo was, with better quality. If your pipeline still references a GPT-3.5 model string, it is running on borrowed time.
GPT-3.5 retirement timeline
| Model | Released | Deprecated / Retired | Status |
|---|---|---|---|
| gpt-3.5-turbo-0301 | Mar 1, 2023 | Jun 13, 2023 | Retired |
| gpt-3.5-turbo-0613 | Jun 13, 2023 | Sep 13, 2024 | Retired |
| text-davinci-003 | Nov 28, 2022 | Jan 4, 2024 | Retired |
| text-davinci-002 | Mar 15, 2022 | Jan 4, 2024 | Retired |
Live status for every OpenAI model — including GPT-4, GPT-4 Turbo, GPT-4o and the o-series — is tracked on our OpenAI model deprecation page, regenerated at every deploy.
What to migrate to
OpenAI's guidance for GPT-3.5 Turbo workloads points to GPT-4o mini: it outperforms GPT-3.5 Turbo on standard benchmarks while costing less per token, and it supports larger context windows. For most chat-completion workloads the migration is a model-string change plus a regression pass on your prompts — GPT-3.5-era prompt phrasing sometimes over-triggers safety behaviors or verbosity in newer models, so test before you flip production.
Migration checklist
- Inventory model strings — grep your codebase and config for
gpt-3.5,davinci, and any dated snapshot pins. - Check your fallback chains — many apps list GPT-3.5 Turbo as the cheap fallback model. A retired fallback is worse than none.
- Re-run evals — token costs, latency, and output style all shift between model generations.
- Pin a dated snapshot of the replacement — and put its deprecation horizon on your calendar, because this cycle repeats.
The bigger pattern: AI models now have EOL dates
Every major provider — OpenAI, Google Gemini, Meta Llama, Mistral — retires models on rolling schedules measured in months, not years. It is the fastest EOL cadence of any software category we track, and it is the one least covered by traditional lifecycle tooling. Our AI model deprecation hub tracks shutdown dates across all four providers in one place.