← all posts
migration·Lukas Babaliauskas·Nov 18, 2026·14 min read

GPT-4o and GPT-4.1 API shutdown on January 14, 2027: dates, replacements, how to migrate

GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini shut down on OpenAI's API on January 14, 2027. The snapshots, the gpt-6-luna replacement, how to migrate.

On OpenAI's API, GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini shut down on January 14, 2027. OpenAI's deprecations page lists five snapshots for that date: gpt-4o-2024-08-06, gpt-4o-2024-11-20, gpt-4o-mini-2024-07-18, gpt-4.1-2025-04-14 and gpt-4.1-mini-2025-04-14. The recommended replacement for all five is gpt-6-luna. OpenAI announced the shutdown on October 14, 2026, 92 days ahead, although its policy promises at least six months for generally available models. Fine-tuned models built on these snapshots shut down the same day.

Changing the model string is not the whole migration:

  • +gpt-6-luna is a reasoning model. Its reasoning effort defaults to medium. At any effort above none, OpenAI says to remove temperature, top_p and top_logprobs, and Chat Completions supports function calling only at none. Reasoning tokens are billed as output and count toward the output cap.
  • +Fine-tunes have no self-serve successor. gpt-6-luna can't be fine-tuned, and from January 6, 2027 existing customers can no longer start new self-serve fine-tuning jobs.
  • +The other dates you've seen are not this one. ChatGPT dropped GPT-4o on February 13, 2026, and Azure OpenAI runs most of these models until April 14, 2027. Neither applies to OpenAI's API.

The rest of this post covers every upcoming GPT-4o and GPT-4.1 date, what happens to the undated gpt-4o and gpt-4.1 names, what changes in the request, and a six-step migration. The facts come from OpenAI's deprecations page ↗, its model and pricing pages, its GPT-6 guide ↗, and Azure's retirement schedule ↗, all read on November 17, 2026. Check the page again before you act on a date.

GPT-4o and GPT-4.1 shutdown dates on OpenAI's API

Model IDShutdown dateAnnouncedListed replacement
gpt-4o-2024-08-06, the snapshot gpt-4o points toJanuary 14, 2027October 14, 2026gpt-6-luna
gpt-4o-2024-11-20January 14, 2027October 14, 2026gpt-6-luna
gpt-4o-mini-2024-07-18, the snapshot gpt-4o-mini points toJanuary 14, 2027October 14, 2026gpt-6-luna
gpt-4.1-2025-04-14, the snapshot gpt-4.1 points toJanuary 14, 2027October 14, 2026gpt-6-luna
gpt-4.1-mini-2025-04-14, the snapshot gpt-4.1-mini points toJanuary 14, 2027October 14, 2026gpt-6-luna
gpt-4o-2024-05-13Shut down October 23, 2026April 22, 2026gpt-5.6-sol
gpt-4.1-nano, gpt-4.1-nano-2025-04-14Shut down October 23, 2026April 22, 2026gpt-5.6-luna
gpt-4o-search-preview-2025-03-11, gpt-4o-mini-search-preview-2025-03-11Shut down July 23, 2026April 22, 2026gpt-5.6-terra
chatgpt-4o-latestShut down February 17, 2026November 18, 2025gpt-5.1-chat-latest

The first five rows are the January 14 shutdown. The rest already happened. They matter only if a fallback list or an old config still names them, because those calls already fail.

Ninety-two days is short by OpenAI's own rules. The policy gives generally available models at least six months unless safety or compliance concerns require a faster timeline, and the deprecations page gives no reason for this one. The previous GPT-4o and GPT-4.1 shutdown, which took gpt-4o-2024-05-13 and gpt-4.1-nano on October 23, was announced 184 days ahead.

The GPT-4o audio, realtime and speech models have their own dates

The gpt-4o name is also on OpenAI's audio models, and they shut down on three other dates:

Shutdown dateModel IDListed replacement
January 6, 2027gpt-4o-mini-tts, gpt-4o-mini-tts-2025-03-20, gpt-4o-mini-tts-2025-12-15gpt-realtime-2.1-mini
January 20, 2027gpt-4o-audio, gpt-4o-mini-audiogpt-audio-1.5
January 20, 2027gpt-4o-realtimegpt-realtime-2.1
January 20, 2027gpt-4o-mini-realtimegpt-realtime-2.1-mini
January 20, 2027gpt-4o-mini-transcribe-2025-03-20gpt-4o-mini-transcribe-2025-12-15
February 26, 2027gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-mini-transcribe-2025-12-15, gpt-4o-transcribe-diarizegpt-transcribe or gpt-live-transcribe

One row is a trap. OpenAI lists gpt-4o-mini-transcribe-2025-12-15 as the replacement for the March 2025 transcribe snapshot, and that replacement shuts down on February 26, 2027, five weeks later. Go straight to gpt-transcribe or gpt-live-transcribe.

What happens to the undated gpt-4o and gpt-4.1 names

Most code calls the short name, not the dated snapshot. OpenAI's model pages say where each one points: gpt-4o to gpt-4o-2024-08-06, gpt-4o-mini to gpt-4o-mini-2024-07-18, gpt-4.1 to gpt-4.1-2025-04-14, and gpt-4.1-mini to gpt-4.1-mini-2025-04-14. Each of those pages marks the model as Deprecated, and every snapshot it lists has either shut down already or shuts down on January 14.

The deprecations page lists each model by name with its dated snapshots. Unlike rows such as gpt-5, it doesn't list the short ID itself, and OpenAI hasn't said whether a call to gpt-4o fails on January 15 or gets pointed somewhere else. Plan for it to fail. If OpenAI did repoint it, your feature would move to a different model on a date you didn't choose, with no test run, which is the worse of the two outcomes.

Every response names the snapshot that answered, so you can confirm what a short name resolves to today:

python
from openai import OpenAI

client = OpenAI()
resp = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "ping"}],
)
print(resp.model)  # the dated snapshot that served the request, e.g. gpt-4o-2024-08-06

What replaces GPT-4o and GPT-4.1: gpt-6-luna

OpenAI lists one replacement for all four models: gpt-6-luna, released on September 22, 2026 and the lowest-priced model in the GPT-6 family. Its model page calls it "our most efficient model for focused, high-volume tasks." On paper, at standard prices per million tokens:

gpt-4ogpt-4o-minigpt-4.1gpt-4.1-minigpt-6-luna
Input$2.50$0.15$2.00$0.40$0.10
Cached input$1.25$0.075$0.50$0.10$0.01
Output$10.00$0.60$8.00$1.60$0.50
Context window128,000128,0001,047,5761,047,5761,050,000, of which 922,000 input
Max output tokens16,38416,38432,76832,768128,000
Knowledge cutoffOctober 1, 2023October 1, 2023June 1, 2024June 1, 2024May 18, 2026
ReasoningNoNoNoNoYes, medium by default
Fine-tuningYesYesYesYesNo

Every per-token price goes down. gpt-6-luna's input price is a twenty-fifth of gpt-4o's and its output price a twentieth. Two things work against that:

  • +Reasoning tokens. At the default medium effort the model reasons before it answers, and OpenAI bills those reasoning tokens as output. OpenAI's reasoning guide says a request can use anywhere from a few hundred to tens of thousands of them.
  • +Price rules the old models don't have. gpt-6-luna charges for cache writes ($0.125 per million tokens for a 30-minute write, 1.25 times its input rate), and a prompt over 272,000 input tokens is billed at twice the input and cache rates and 1.5 times the output rate for the whole request. The GPT-4.1 models have a single rate at every prompt length, so a GPT-4.1 route that used its million-token window pays the long-context rate on gpt-6-luna. A prompt over 922,000 tokens doesn't fit at all.

So the per-request cost depends on the reasoning effort you pick and on your traffic, and step 5 is where you measure it.

Whether gpt-6-luna is enough is a question for your own traffic. OpenAI recommends its lowest-priced GPT-6 model for all four, including GPT-4o, which its model page still calls a flagship. If it fails your comparison on a gpt-4o or gpt-4.1 route, the next tiers up are gpt-6.1-terra ($1 input, $5 output), which supports none, and gpt-6.1-sol ($2 input, $10 output), which doesn't. Both need the Responses API for any tool calling. Avoid migrating onto a model with its own date: gpt-5 shuts down on December 11, 2026 and gpt-5.1 on April 1, 2027.

Fine-tuned GPT-4o and GPT-4.1 models

A fine-tuned model follows its base model's shutdown date. Every model that starts with ft:gpt-4o-2024-08-06, ft:gpt-4o-mini-2024-07-18, ft:gpt-4.1-2025-04-14 or ft:gpt-4.1-mini-2025-04-14 stops serving on January 14, 2027. This is the hardest part of the migration:

  • +gpt-6-luna's model page lists fine-tuning as not supported.
  • +OpenAI has been closing self-serve fine-tuning in stages. Since May 7, 2026, organizations that never ran fine-tuning can't start. Since July 2, 2026, organizations that haven't run inference on a fine-tuned model in the past 60 days can't create jobs. From January 6, 2027, active existing customers can't create new jobs either.
  • +After January 14 the fine-tune is gone, and with it the only record of how it behaved beyond whatever you logged.

The realistic path is to replace the fine-tune with a prompt: few-shot examples drawn from the training set, plus Structured Outputs where the fine-tune enforced a format. Then measure the prompted gpt-6-luna against the fine-tune's own outputs on held-out cases while the fine-tune still answers. OpenAI also says some developers may be able to provision dedicated capacity to keep using a model after its shutdown date, through its sales team.

Fine-tunes trained on Azure OpenAI last longer. Azure lists the deployment retirement for fine-tuned gpt-4o and gpt-4o-mini on October 1, 2027, and for the fine-tuned gpt-4.1 models on October 14, 2027.

Azure OpenAI retires GPT-4o and GPT-4.1 later

Azure OpenAI, documented under Microsoft Foundry, sets its own retirement schedule ↗:

Model and versionOpenAI APIAzure OpenAIAzure's listed replacement
gpt-4o 2024-05-13Shut down October 23, 2026January 14, 2027gpt-5.1 2025-11-13
gpt-4o 2024-08-06 and 2024-11-20January 14, 2027April 14, 2027gpt-5.1
gpt-4o-mini 2024-07-18January 14, 2027April 14, 2027None listed
gpt-4.1 and gpt-4.1-mini 2025-04-14January 14, 2027April 14, 2027None listed
gpt-4.1-nano 2025-04-14Shut down October 23, 2026April 14, 2027None listed

For gpt-4o 2024-05-13, Azure says Global Standard and Data Zone Standard deployments upgrade to gpt-5.1 automatically at retirement, and Provisioned deployments don't: they need replacement deployments before January 14. Regional Standard deployments that Azure already retired stay retired. gpt-5.1 itself retires on Azure on July 13, 2027. Azure's page was last updated on September 23, 2026, three weeks before OpenAI's October 14 notice, so check it again before you plan around an April date.

ChatGPT retired GPT-4o in February 2026, the API is a different date

Many "GPT-4o retirement" headlines are about ChatGPT. On January 29, 2026, OpenAI announced ↗ that GPT-4o, GPT-4.1, GPT-4.1 mini and o4-mini would leave ChatGPT on February 13, 2026, and said that in the API there were "no changes at this time." OpenAI also said that only 0.1% of users were still choosing GPT-4o each day. Its help center ↗ confirms the February 13 retirement, with one exception: ChatGPT Business, Enterprise and Edu customers kept GPT-4o in Custom GPTs until April 3, 2026. The API notice came eight and a half months after the announcement, on October 14, 2026.

A model leaving a chat app tells you nothing about its API date. For the API, OpenAI's deprecations page is the schedule that counts.

Step 1: Find every GPT-4o and GPT-4.1 call

Model strings hide in more places than application code: environment files, Terraform, prompt registries, notebooks, batch jobs, and the eval harness, where a judge model is easy to forget. Start with a grep, then read every hit (drop -h to see the files):

bash
grep -rhoE "(ft:)?gpt-4(o|\.1)[a-z0-9.:_-]*" \
  --exclude-dir={node_modules,.git,.venv,dist} . | sort | uniq -c | sort -rn

Then ask the API. GET /v1/models lists every model your organization can call, fine-tunes included, and OpenAI documents an optional shutdown_date field on each:

python
from openai import OpenAI

for m in OpenAI().models.list():
    if m.id.startswith(("gpt-4o", "gpt-4.1", "ft:gpt-4o", "ft:gpt-4.1")):
        print(m.id, getattr(m, "shutdown_date", None))

For each ID, note the route that calls it and what the route does with it: tool calls, JSON output, free text or audio. That decides how much testing step 5 needs.

Step 2: Record a baseline while GPT-4o still answers

Once a model shuts down, nobody can record what it did. Before January 14, or January 6 for gpt-4o-mini-tts, capture 100 to 300 real production inputs per route, stratified by route, with what production actually did: the tools called, their arguments, the JSON fields, the refusals. For a fine-tune this recording is the only specification of its behavior that will survive it.

The method is in Build a golden dataset for LLM evals from production traffic.

Step 3: Pick the target model and reasoning effort per route

Start each route on gpt-6-luna at reasoning effort none, which is the smallest change from a non-reasoning model. Run the same cases at low as well, because the reasoning effort is now a setting you choose, and it changes both quality and cost. Move a route up to gpt-6.1-terra only if both fail on that route, and through the Responses API if the route calls tools. Write down the reason for each choice, because the next deprecation will ask the same question.

Step 4: Update the request for gpt-6-luna

OpenAI's GPT-6 migration guide ↗ lists the compatibility checks. From a GPT-4o or GPT-4.1 request, these are the ones that bite:

In your GPT-4o or GPT-4.1 requestOn gpt-6-lunaWhat to do
No reasoning settingReasoning effort defaults to mediumSet it explicitly: reasoning_effort in Chat Completions, reasoning.effort in Responses
temperature, top_p, top_logprobs, and logprobs in Chat CompletionsMust be removed when reasoning effort isn't noneRemove them, or keep effort at none
tools in Chat CompletionsFunction calling works only with reasoning_effort: "none"Use none, or move tool-calling routes to the Responses API
An output cap sized for the visible answerReasoning tokens count toward it. In the Responses API, a response that reaches it ends with status incomplete, possibly before any visible textRaise the cap. OpenAI suggests reserving at least 25,000 tokens for reasoning and output when you start
prompt_cache_retentionReplaced by prompt_cache_options.ttl set to "30m"Switch the parameter, and budget for cache writes
prediction (Predicted Outputs)Not among gpt-6-luna's listed featuresRemove it and recheck latency

The closest drop-in for a Chat Completions route with tools keeps reasoning off:

python
# Before
resp = client.chat.completions.create(
    model="gpt-4o",
    messages=messages,
    tools=tools,
    temperature=0,
)

# After: reasoning off, so Chat Completions function calling works and temperature is allowed
resp = client.chat.completions.create(
    model="gpt-6-luna",
    messages=messages,
    tools=tools,
    reasoning_effort="none",
    temperature=0,
)

That is the smallest change, not necessarily the best one. To try a tool-calling route with reasoning on, move it to the Responses API, where tools work at any effort, and compare both versions in step 5.

Step 5: Replay real traffic on both models and diff per route

Everything so far makes gpt-6-luna answer. This step finds out whether it still does the same job. The regressions that matter in a model swap mostly read fine: an argument changes format, an enum field moves one notch, a request the old model refused becomes a tool call.

  • +Assert deterministically first. Check the tools called, the argument formats and every structured field. Use an LLM judge only for free text.
  • +Sample each case several times when reasoning is on. Without temperature, one sample can't tell a regression from sampling noise. Score the share of samples that pass.
  • +Measure cost and latency per request. The usage object reports reasoning tokens (completion_tokens_details.reasoning_tokens in Chat Completions, output_tokens_details.reasoning_tokens in Responses), which is where a cheaper list price can turn into a bigger bill.
  • +Compare per route, never one average. A drop on one route disappears in an overall score.

The full walkthrough, with code and the six regressions it caught on a real model upgrade, is in How to catch LLM regressions before a model upgrade ships.

Step 6: Cut over before January 14, with a rollback path

Switch one route at a time behind a flag, starting with the route the diff cleared most clearly. Keep the old snapshot as the fallback while it still answers, then take it out of every fallback list before January 14: a fallback to a shut-down model fails exactly when you need it. Watch error rates, cost, latency and response.model per route for a week after each switch.

gpt-6-luna has no shutdown date yet. Put a check for it in CI rather than on a calendar: the deprecation schedule post has a short script that fails the build when a pinned model's shutdown_date is less than 90 days away or the model is no longer listed.

Doing this with EvalShift

Disclosure: I'm the founder of EvalShift, and steps 2 and 5 are the part it automates. The capture SDK wraps the OpenAI client and records real runs, including tool calls, and compare replays them against the replacement:

python
from openai import OpenAI
from evalshift import capture
from evalshift.adapters.openai import wrap_openai

client = wrap_openai(OpenAI())  # each create() inside a capture records one model call

TOOLS = [...]  # the tools the agent is offered, or [] for none

@capture.agent(suite="support", redact=True, tools=TOOLS)  # no-op unless EVALSHIFT_CAPTURE=1
def handle_ticket(message): ...
bash
pip install evalshift
EVALSHIFT_CAPTURE=1 python my_agent.py   # record production-shaped runs
evalshift capture sync                   # recorded runs -> golden suite
evalshift compare --suite-name support --to gpt-6-luna

Both arms are live calls, so run it before January 14. EvalShift's config has no reasoning-effort setting, so compare measures gpt-6-luna at OpenAI's default, medium: the change you get by editing only the model string, not the none setting from step 3. EvalShift sends temperature=0 on every call. If a model rejects it, EvalShift resends without it and marks the run as non-deterministic in the report, so set defaults.samples_per_example above 1 for the reasoning model. The max_cost_increase threshold in the migration_policy fails the verdict if the paired cost rises by more than you allow, which is where reasoning tokens outweighing the lower price would show up. Docs: /docs/sdk-adapters, /docs/captures, /docs/migration-policy.

FAQ

When is GPT-4o being shut down? On OpenAI's API, the gpt-4o-2024-08-06 and gpt-4o-2024-11-20 snapshots shut down on January 14, 2027, together with gpt-4o-mini-2024-07-18. OpenAI announced it on October 14, 2026. The oldest snapshot, gpt-4o-2024-05-13, already shut down on October 23, 2026, and ChatGPT stopped offering GPT-4o on February 13, 2026.

Is GPT-4.1 being deprecated? Yes. gpt-4.1-2025-04-14 and gpt-4.1-mini-2025-04-14 shut down on OpenAI's API on January 14, 2027, and OpenAI's model pages mark gpt-4.1 and gpt-4.1-mini as deprecated. gpt-4.1-nano already shut down on October 23, 2026. On Azure OpenAI the GPT-4.1 models retire on April 14, 2027.

What replaces GPT-4o and GPT-4o mini? OpenAI lists gpt-6-luna as the replacement for GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini. It is a reasoning model with a lower per-token price than any of them, it defaults to medium reasoning effort, and it supports none, which is the closest match to the old non-reasoning behavior. If it falls short on a route, gpt-6.1-terra is the next tier up, through the Responses API for tool calls.

Will the gpt-4o model name keep working after January 14, 2027? OpenAI hasn't said. gpt-4o points to gpt-4o-2024-08-06, OpenAI's model page marks it as deprecated, and every snapshot it lists is shut down or shuts down on January 14. Plan for calls to gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini to fail.

What happens to fine-tuned GPT-4o and GPT-4.1 models? They shut down with their base model on January 14, 2027. gpt-6-luna can't be fine-tuned, and from January 6, 2027 existing customers can no longer create new self-serve fine-tuning jobs, so the usual path is to replace the fine-tune with a few-shot prompt and test it against the fine-tune's outputs before the date.

Is GPT-4o still available in ChatGPT? No. OpenAI retired GPT-4o, GPT-4.1, GPT-4.1 mini and o4-mini from ChatGPT on February 13, 2026, and Business, Enterprise and Edu customers kept GPT-4o in Custom GPTs only until April 3, 2026. That retirement didn't change the API, where the shutdown date is January 14, 2027.

How long can I use GPT-4o on Azure OpenAI? Azure lists April 14, 2027 for gpt-4o versions 2024-08-06 and 2024-11-20, gpt-4o-mini and the GPT-4.1 models, and January 14, 2027 for gpt-4o 2024-05-13. Fine-tuned deployments last until October 2027.

Keep reading