OpenAI has switched off GPT-4 and o4-mini in its API. Some swaps cost more

The list covers gpt-3.5-turbo, gpt-4, gpt-4-turbo, gpt-4-1106-preview, the first gpt-4o snapshot (gpt-4o-2024-05-13), gpt-4.1-nano, o1, o1-pro, o3-mini, o4-mini and the image model gpt-image-1. Fine-tuned versions went with their base models. APIYI, a developer API provider, counted 12 model entries and five fine-tuned model categories in the shutdown.

“The model OpenAI suggests isn't always the one you should pick.”

OpenAI names a replacement for each. The text models all point to its GPT-5.6 family: gpt-5.6-sol for the GPT-4, gpt-4o, o1 and o3-mini models, gpt-5.6-terra for gpt-3.5-turbo and o4-mini, and gpt-5.6-luna for gpt-4.1-nano.

Why it matters

If you added AI to a product or a client project in 2023 or 2024, there's a fair chance it names one of these models: the support bot, the lead-scoring script, the "summarise this call" step in an automation. Builds like that often go untouched for years, because they work.

The recommended swaps are not like-for-like on price. o4-mini cost $1.10 per million input tokens and $4.40 per million output. Its named replacement, gpt-5.6-terra, is $2 and $12. gpt-3.5-turbo was $0.50 and $1.50, and OpenAI points that to the same terra model. gpt-4.1-nano's swap, gpt-5.6-luna, doubles input and triples output.

The model OpenAI suggests isn't always the one you should pick. GPT-6 Luna is $0.10 in and $0.50 out, cheaper than any of the text replacements named for this batch, and it's what OpenAI names for GPT-4o and GPT-4.1 in January. The catch: it defaults to medium reasoning effort, and in Chat Completions it only supports function calling with reasoning effort set to none. Cheaper on paper can mean a weekend of rework.

If you run an agency, the risk is client builds. A bot you shipped two years ago may have been failing since Friday, and the client may not have noticed yet. Better they hear it from you.

If you sell SaaS, check every feature that names a model in code, not just the main one. And note the dry joke on o4-mini's model page: it says the model was "succeeded by GPT-5 Mini", which itself shuts down on 11 December.

What to do this week

1. Search your code, your automation tools and every client build for these model names: gpt-3.5-turbo, gpt-4, gpt-4-turbo, gpt-4o-2024-05-13, gpt-4.1-nano, o1, o3-mini, o4-mini and gpt-image-1. Then check your error logs from 23 October onwards.
2. Price the swap before you make it. Take a feature handling 20 million input and 4 million output tokens a month. On o4-mini that cost $39.60. On gpt-5.6-terra it's $88. On gpt-5.6-luna it's $8.80, and on GPT-6 Luna it's $4 (FER calculations from OpenAI's list prices).
3. Test the cheapest option first. Run 50 real inputs through it, compare the answers with your old output, and only move up a tier if it fails.
4. Put the next dates in the diary. Agent Builder, Evals and reusable prompts in the API go on 30 November. Older image models, including gpt-image-1-mini and gpt-image-1.5, go on 1 December. GPT-5, GPT-5 mini, GPT-5 nano, GPT-5 pro, o3 and o3-pro go on 11 December. GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini go on 14 January.
5. Stop finding out by outage. OpenAI's List models endpoint returns a shutdown_date field for each model, so a weekly check can flag anything you call that has a date on it.

The 14 January batch was announced on 14 October: three months' notice, when OpenAI's own policy promises at least six for generally available models unless safety or compliance needs a faster timeline. Don't count on the long warning next time. Check this week, and plan the GPT-4o move before Christmas eats the time.