GPT-4o API Retirement: How to Migrate Before January 2027

Last updated: 22 October 2026

If your chatbot, automation, plugin or SaaS feature calls GPT-4o or GPT-4.1 through the OpenAI API, it will stop working on 14 January 2027 unless you switch models. That gives developers about three months, and some older models, such as the first GPT-4o snapshot and GPT-4.1 nano, shut down even sooner, on 23 October 2026.

This guide covers exactly which models are retiring, what to switch to, how the costs compare, the technical changes that break code during migration, and a step-by-step plan to finish before the holidays.

Which OpenAI models are being retired?

OpenAI is retiring the GPT-4o and GPT-4.1 families from its API in two waves: older snapshots on 23 October 2026 and the main snapshots on 14 January 2027. Fine-tuned models follow their base model’s date.

OpenAI keeps the full schedule on its official deprecations page. The ChatGPT app already removed GPT-4o and GPT-4.1 on 13 February 2026, so this change affects only API-built products. For what replaced them in the app, see ChatGPT’s new features in 2026.

What should you replace GPT-4o with?

OpenAI recommends GPT-6 Luna as the default replacement for GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini. GPT-6 Luna is the smallest model in OpenAI’s GPT-6 family, with a context window of 1,050,000 tokens and a knowledge cutoff of 18 May 2026.

For harder work such as complex reasoning or long agent tasks, consider GPT-6.1 Sol ($2.00 input, $10.00 output) instead. Our guide to GPT-6 models compares Astra, Sol and Luna, and model routing shows how to send easy requests to Luna and hard ones to Sol.

How much cheaper is GPT-6 Luna than GPT-4o?

GPT-6 Luna is about 25 times cheaper than GPT-4o on input and 20 times cheaper on output at list prices. It is also cheaper than GPT-4o mini, which many teams used to save money.

Worked example: a support chatbot

Take a chatbot handling 20,000 conversations a month, each using about 1,500 input tokens and 500 output tokens. That is 30 million input and 10 million output tokens a month.

• GPT-4o: 30 x $2.50 = $75 input plus 10 x $10 = $100 output, so $175 a month.
• GPT-4.1: 30 x $2.00 = $60 plus 10 x $8 = $80, so $140 a month.
• GPT-4o mini: 30 x $0.15 = $4.50 plus 10 x $0.60 = $6, so $10.50 a month.
• GPT-6 Luna (reasoning off): 30 x $0.10 = $3 plus 10 x $0.50 = $5, so $8 a month.

What breaks when you switch to GPT-6 Luna?

Changing the model name is the easy part; the real breaking changes are reasoning defaults, function calling rules and unsupported APIs. Test these four areas before you switch production traffic.

1. Reasoning is on by default. GPT-6 Luna reasons at medium effort unless you change it. OpenAI’s migration guide suggests starting with reasoning set to none, then raising it only where quality needs it.
2. Function calling in Chat Completions needs reasoning off. In the Chat Completions API, Luna supports function calling only with reasoning_effort set to none. To combine reasoning with tools, move to the Responses API.
3. Sampling settings disappear with reasoning. When reasoning is enabled, temperature, top_p, logprobs and top_logprobs are not available, which GPT-4.1 supported. Code that passes them may error.
4. Assistants and Realtime. Reports on the migration say Luna does not run on the Assistants and Realtime APIs, so systems built on them need rebuilding on the Responses API or a current realtime model.

Also re-check prompts that relied on GPT-4o’s tone or formatting, and supply facts in the prompt rather than relying on the model’s memory. Pairing Luna with retrieval-augmented generation is the safest way to keep answers accurate.

What happens to fine-tuned GPT-4o models?

Fine-tuned GPT-4o and GPT-4.1 models stop working on the same day as their base model, 14 January 2027, and GPT-6 Luna cannot be fine-tuned. OpenAI is also winding down self-serve fine-tuning: active customers can start new training jobs only until 6 January 2027.

You have three realistic options:

• Prompting with examples: move the behavior into a system prompt with 5 to 20 high-quality examples. With prompt caching, long instructions cost little ($0.01 per million cached tokens on Luna).
• Retrieval: if the fine-tune taught product knowledge, store that knowledge in a search index and pass the relevant pieces into each request.
• Another provider: some providers still offer fine-tuning or open-weight models you can tune yourself; compare options in our LLM API pricing comparison and open-source AI models guide.

How to migrate from GPT-4o step by step

The safest migration takes four to six weeks: inventory, baseline, test, stage and switch. Aim to finish by mid-December 2026 so you have buffer before the 14 January 2027 shutdown.

1. Inventory every model ID. Search code, environment variables, no-code tools (Zapier, Make, n8n) and vendor products for gpt-4o, gpt-4o-mini, gpt-4.1 and fine-tune IDs. Check the Usage page in the OpenAI dashboard to see which models still get traffic.
2. Ask your vendors. If a SaaS tool you use is built on GPT-4o, ask which model it will move to and when.
3. Record a baseline. Save 100 to 200 real requests with current outputs, latency and cost.
4. Test GPT-6 Luna with reasoning off. Compare quality on your baseline set, then try low reasoning only for the cases that fail.
5. Fix breaking changes. Remove unsupported sampling parameters, move tool-plus-reasoning flows to the Responses API, and rebuild Assistants or Realtime flows.
6. Stage the rollout. Send 10 percent of traffic to Luna, watch errors and complaints, then increase. Store the model name in configuration so the next migration is a one-line change.

Azure OpenAI and other platforms

Azure OpenAI customers get more time: Microsoft lists 14 April 2027 as the retirement date for the same GPT-4o and GPT-4.1 models, and names GPT-5.1 as its suggested replacement for GPT-4o while GPT-6 Luna is not yet listed on Azure. If you use OpenAI models through another platform, check that platform’s own deprecation notices, since dates can differ from OpenAI’s.

Teams with EU customers should note that GPT-6 Luna supports EU data residency, but OpenAI adds a 10 percent surcharge for regional processing on models released on or after 5 March 2026. If you are reconsidering providers entirely, compare Claude API pricing and Gemini API pricing.

Frequently asked questions

Your next step

Run a search for “gpt-4o” and “gpt-4.1” across your code and tools today, then test GPT-6 Luna with reasoning off on a small set of real requests this week. The earlier you start, the more time you have to fix edge cases and enjoy the lower bill. Dates and prices can change, so confirm everything on OpenAI’s deprecations and pricing pages before you plan your rollout.

Sources

• Deprecations, OpenAI API docs, checked 22 October 2026
• Using the latest model (migration guide), OpenAI API docs, checked 22 October 2026
• GPT-6 Luna model page, OpenAI API docs, checked 22 October 2026
• OpenAI GPT-4o API shutdown, The New Stack, 15 October 2026, checked 22 October 2026
• GPT-4o and GPT-4.1 leave OpenAI’s API on 14 January 2027, DataNorth AI, checked 22 October 2026