GPT-4o and GPT-4.1 Leave OpenAI's API on January 14, 2027

On October 14, 2026, OpenAI added a new entry to the top of its API deprecations page: five GPT-4 series snapshots will be removed from the API on January 14, 2027. The list is GPT-4o in its August and November 2024 versions, GPT-4o mini, GPT-4.1 and GPT-4.1 mini. OpenAI names one recommended replacement for all of them, gpt-6-luna, and says fine-tuned versions follow the same date as their base model.

That is three months' warning for what were, until GPT-5 arrived in August 2025, common default models in chatbot builders, automation platforms and customer-service integrations sold to small and mid-sized businesses. Many of those integrations were set up once and never touched again. If your company has a support bot, a lead-qualification flow or a content pipeline that was built in 2024 or early 2025, there is a fair chance it still calls one of these IDs and nobody on the team remembers which.

This guide is about the developer API only. ChatGPT's own removal of GPT-4o from the model picker happened on February 13, 2026 and is covered in our ChatGPT GPT-4o retirement guide. Here the question is narrower and more expensive to get wrong: which systems break on January 14, what they should move to, and what the move does to your bill and your output.

OpenAI publishes every API retirement on one page, newest at the top. The October 14 entry is short. It says developers using certain GPT-4 series models were notified that day of their deprecation and removal from the API on January 14, 2027, then lists the affected snapshots and a replacement for each. We fetched OpenAI's deprecations page on October 15, 2026, and the table below is a direct transcription of that entry.

Two things on that page matter as much as the table. First, OpenAI says it notifies customers who are actively using a model by email. If nobody in your company received that email, it may be because the API key belongs to an agency, a freelancer or a tool vendor rather than to you. Second, the page explains that a model can be queried for its own shutdown date: the List models endpoint returns an optional shutdown_date field as a YYYY-MM-DD string, or null when no date has been announced. That one field is the cheapest audit you can run, and we come back to it in section 03.

The failure case here is the obvious one. A team reads "GPT-4o is being retired," remembers that ChatGPT already dropped it in February, and files the news as old. The February change touched the consumer app. This one touches the API, and the API is where the quiet, unattended integrations live.

OpenAI lists dated snapshot IDs, not the short aliases most code actually uses. Many integrations call gpt-4o or gpt-4.1 with no date, and OpenAI resolves the alias to a current snapshot. The October 14 entry does not say what happens to those bare aliases on January 14. It names the snapshots, and every current snapshot behind the aliases is on the list.

We are not going to guess at alias behaviour after the date, because the page does not state it and OpenAI's own definitions on the same page draw a hard line: at shutdown, a model is "no longer accessible." The safe reading is that anything resolving to a listed snapshot stops working on January 14 unless OpenAI points the alias somewhere else, and nothing on the page says it will. Treat an alias as a listed snapshot wearing a short name.

There is also an earlier date to know about. An April 22, 2026 notice on the same page already scheduled a batch of legacy snapshots for October 23, 2026, and two of them belong to these families: gpt-4o-2024-05-13, the original GPT-4o, and gpt-4.1-nano. Their listed substitutes are gpt-5.6-sol and gpt-5.6-luna, not gpt-6-luna, because that notice predates the GPT-6 line. If you move everything at once, use the newer recommendation. Our model alias and retirement ledger tracks the full set of OpenAI, Anthropic and Google aliases and which snapshot each one points to.

Most of the businesses this affects did not write code against the OpenAI API. They picked a model from a dropdown in a chatbot builder, a workflow tool or a CRM plug-in, sometime in 2024, and the dropdown has shown the same value ever since. Finding every instance is an inventory job, not an engineering one, and it has three layers.

The first layer is your own OpenAI account. The usage dashboard breaks spend down by model, so a GPT-4o or GPT-4.1 line in the current month means something is still calling it. The List models endpoint gives the same answer in code: for each model you use, read its shutdown_date and alert on anything that is not null. That check is a few lines, runs in seconds, and catches every future notice as well as this one.

The second layer is every third-party tool that holds its own key. Automation platforms, chatbot builders, email tools and AI writing assistants often let you choose a model per workflow, and the default they offered in 2024 was usually GPT-4o or GPT-4o mini. Open each one, find the model setting, and write down what it says. If the vendor supplies the key and hides the model, ask the vendor in writing which model your plan uses and what their migration date is.

The third layer is code. Search every repository and configuration store for the strings gpt-4o and gpt-4.1, including environment files and secrets managers, where model names are often stored as settings. The failure case is the one everybody hits: the search finds the obvious call in the application, misses the model name sitting in a scheduled job's environment variable, and the job fails at 3 a.m. on January 15.

On token price alone, this migration saves money for everyone. We took the four retiring models' list prices from their individual OpenAI model pages and gpt-6-luna's from its page and OpenAI's pricing table, all read on October 15, 2026. The per-model pages and the standard-tier rows on OpenAI's pricing page agree on every figure below.

The ratios are large. GPT-4o to gpt-6-luna is 25× cheaper on input and 20× on output. GPT-4.1 is 20× and 16×. Even the mini models get cheaper: GPT-4o mini drops by a third on input and a sixth on output, and GPT-4.1 mini by three quarters and more than two thirds. A workload of 10 million input and 2 million output tokens a month costs $45 on GPT-4o and $2 on gpt-6-luna at these list prices.

The catch is in the word "output." gpt-6-luna is a reasoning model. At its default medium effort it produces reasoning tokens before the answer, and those are billed as output. A prompt that returned 300 tokens from GPT-4o may return 300 tokens plus a few hundred tokens of reasoning from gpt-6-luna. At $0.50 per million that is still far cheaper than GPT-4o's $10, but the per-request token count will not match your old logs, and a cost alert set on tokens rather than dollars will fire. Set reasoning effort to none for simple classification and extraction tasks and the comparison is like for like. For the wider price picture across labs, our per-million-token price index carries every current rate.

GPT-4o and GPT-4.1 answer directly. gpt-6-luna thinks first. OpenAI's model page lists reasoning effort levels of none, low, medium, high, xhigh and max, with medium as the default, and adds a constraint that matters to almost every chatbot built in 2024: Chat Completions supports function calling only when reasoning effort is set to none. Built-in tools and function calling at any other effort level require the Responses API.

For a plain question-and-answer bot, that means one extra parameter in the request. For an automation that calls tools from Chat Completions, it means choosing between two migrations: keep Chat Completions and set effort to none, which gives you a fast, non-reasoning model roughly in the GPT-4o mould, or move to the Responses API and keep reasoning on. The second is more work and usually the better product. The first is a one-line change that keeps the lights on by January.

The other differences are easy to list and easy to forget. The context window grows from 128,000 tokens on GPT-4o to 1,050,000, with a 922,000-token cap on input. Maximum output rises from 16,384 tokens to 128,000. The knowledge cutoff moves from October 2023 for GPT-4o, and June 2024 for GPT-4.1, to May 18, 2026. A bot whose system prompt pastes in a page of company facts "because the model doesn't know about 2025" can usually drop that padding. A bot that relied on the model not knowing recent events, for a consistent persona or a fixed-date test, will behave differently.

A failure case to look for: a summariser tuned to GPT-4o's 16,384-token output cap often carries a client-side truncation that nobody remembers writing. On a model that can produce 128,000 tokens, that truncation silently cuts long documents short. Audit the limits you set on the old model, not only the ones the new model sets.

A model swap is a product change dressed as a configuration change. The three things that break most often are output format, tool calls and latency, in that order, and each has a cheap test.

Output format. Take fifty real prompts from your logs, run them through both models, and diff the shape of the answers: length, whether JSON stays valid, whether a required field is still present, whether the tone matches the persona in your system prompt. Reasoning models tend to answer more completely and sometimes more verbosely. If a downstream step parses the reply, the parser is the thing under test.

Tool calls. If the integration uses function calling, run every tool through the new model once with effort set to none on Chat Completions, and once on the Responses API with the default effort. Check that argument names and types still match the schema you declared. A model that reasons about a tool before calling it can also decide not to call it, which is correct behaviour and a surprise to a workflow that expected a call every time.

Latency. Reasoning adds time before the first token. For a chat widget that streams, measure time to first token on the new model at effort none and at medium, and pick the setting that keeps the widget feeling responsive. For a batch job, latency rarely matters and medium effort is the better default. The failure case to avoid is a support bot migrated with reasoning on that starts taking seconds to begin each answer, with the drop in live-chat ratings only connected to the model change weeks later.

The GPT-4 notice is the headline because of how many businesses it reaches, but the deprecations page carries several other upcoming shutdowns between now and April 2027. The table below covers the model notices most likely to sit beside a GPT-4o integration; the page also lists GPT Image models on December 1, 2026 and the Evals platform, Agent Builder and reusable prompts on November 30, 2026. A team that is already auditing model names should check these in the same pass rather than coming back for each one.

Read as a set, these dates say something about where OpenAI is steering developers. Every text-model replacement is a GPT-5.6 or GPT-6 model, and text-to-speech is moving to the Realtime API. A company that moves its GPT-4o chat traffic to gpt-6-luna in November and leaves a gpt-4o-transcribe call in a meeting-notes tool will be back here in February. Do the audit once.

The same page that carries the notice also carries OpenAI's policy on how much notice it gives. For generally available models the stated minimum is at least six months. Specialised variants such as chat, Codex and deep-research models get at least three. Preview models can go with much shorter notice, such as two weeks. The policy adds that safety or compliance concerns can shorten any of these, with as much notice as reasonably possible.

The GPT-4 series notice gives three months: October 14 to January 14. GPT-4o and GPT-4.1 were generally available models by any ordinary reading. OpenAI's page does not say which category it applied or why, and we are not going to supply a reason it did not give. What we can say is that the page defines "deprecated" as taking effect the moment a notice is published, and "shut down" as the moment a model stops responding, and that both of those are now dated.

Two practical consequences follow. First, do not plan on an extension. Nothing in OpenAI's text promises one, and the April notice for the older snapshots has held its October 23 date for six months. Second, if three months is not enough for a regulated or contractually bound system, the page says developers may in some cases provision dedicated capacity for continued access after a shutdown date, through OpenAI's sales team. That is an exception to ask about now, not in January.

For fine-tuned models there is one more date. OpenAI's fine-tuning table on the same page says that on January 6, 2027 active existing customers can no longer create new fine-tuning jobs, and that inference on an existing fine-tune ends only when its base model is deprecated. A GPT-4o fine-tune therefore stops on January 14 with its base, and there will be no way to retrain it on the old base after January 6. If a fine-tune carries business logic you cannot reproduce with a prompt, that retraining window is the real deadline, and it is eight days earlier than the headline one. Our AI transformation team runs these inventories and migrations for companies whose integrations were built by someone who has since moved on.