GPT-4o, GPT-4o mini & GPT-4.1 Shut Down on 2027-01-14
Last verified: 2026-10-26 — checked against the OpenAI API Deprecations page. This page covers the 2027-01-14 retirement batch (announced 2026-10-14) and lists the other batches scheduled around it.
Direct answer
On January 14, 2027, OpenAI removes gpt-4o (gpt-4o-2024-08-06, gpt-4o-2024-11-20), gpt-4o-mini (gpt-4o-mini-2024-07-18), gpt-4.1 (gpt-4.1-2025-04-14) and gpt-4.1-mini (gpt-4.1-mini-2025-04-14) from the API. The official replacement for all of them is gpt-6-luna. Fine-tuned versions of these models shut down on the same date. What to do: find every place these IDs are pinned, repoint them to gpt-6-luna (or another current model you have evaluated), re-run your evals, and ship behind a canary before the date.
Who this affects
gpt-4o and gpt-4o-mini have been the default model IDs in countless tutorials, SDK examples, no-code connectors and internal tools since 2024, so the blast radius of this batch is wider than any earlier GPT-5-era retirement. You are affected if any of the following still references one of the retired IDs:
- application code, environment variables, feature flags, CI/CD and infrastructure-as-code;
- SDK or framework defaults (several agent frameworks shipped
gpt-4o-minias the default model); - fine-tuned models built on these bases — they retire on the same day as the base model;
- third-party tools where you typed a model name into a settings field (automation platforms, chat UIs, plugins).
Since the exact dated snapshots are listed, an undated alias such as gpt-4o that resolves to one of them stops working too.
Retirement table (2027-01-14)
Every date and replacement below is taken from the OpenAI API Deprecations page, 2026-10-14 announcement.
| Retired model (shuts down 2027-01-14) | Snapshots named by OpenAI | Official replacement |
|---|---|---|
| GPT-4o | gpt-4o-2024-08-06, gpt-4o-2024-11-20 | gpt-6-luna |
| GPT-4o mini | gpt-4o-mini-2024-07-18 | gpt-6-luna |
| GPT-4.1 | gpt-4.1-2025-04-14 | gpt-6-luna |
| GPT-4.1 mini | gpt-4.1-mini-2025-04-14 | gpt-6-luna |
| Fine-tunes of the above | — | Same schedule as the base model |
gpt-4o → gpt-6-luna: what is not a drop-in
gpt-6-luna is a different model generation, not a renamed GPT-4o. Treat it as a new model and verify the points below against your own workload. Specs are quoted from OpenAI's model pages (gpt-4o page and gpt-6 family pages, checked 2026-10-26).
gpt-4o (retiring) | gpt-6-luna (replacement) | |
|---|---|---|
| OpenAI's positioning | "Fast, intelligent, flexible GPT model" | "Most efficient model for focused, high-volume tasks" |
| Context window | 128,000 tokens | 1,050,000 tokens |
| Max output | 16,384 tokens | 128,000 tokens |
| Knowledge cutoff | Oct 1, 2023 | May 18, 2026 |
| Input / output modalities | Text + image in, text out | Text + image in, text out |
| OpenAI list price (standard, per 1M tokens) | $2.50 input / $10.00 output | $0.10 input / $0.50 output; cached input $0.01 |
Things that commonly need more than a string swap:
- Prompts tuned for a 2024-era model. Instruction-following, verbosity and refusal behaviour differ across generations. Re-run your golden set; do not assume the same system prompt produces the same output distribution.
- Structured Outputs and tool calling. Both are supported on current models, but schema edge cases (optional fields, enums, nested arrays) and parallel tool-call behaviour should be re-tested rather than assumed.
- Token budgets. A far larger context window is a capability, not a cost plan. Anything that silently stuffs more context into each request will raise your bill even at a lower per-token rate. Set explicit limits.
- Prompt caching. OpenAI's current pricing distinguishes cached input from uncached input and, on current models, charges for cache writes (30-minute and 12-hour tiers) as well. Check the pricing page for the exact figures and verify your cache hit rate after cutover instead of assuming it carries over.
- Vision inputs. Both models accept text and image input per OpenAI's model pages, but image-token accounting and detail-level behaviour can differ between generations. Re-check costs on a sample of your real images before moving multimodal traffic.
- Fine-tunes. A fine-tune of
gpt-4ocannot be moved; it has to be re-trained on a supported base model. Budget time for data export and a new training run.
Find every pinned ID
1. Grep your repositories and config
grep -rnE --exclude-dir=node_modules --exclude-dir=.git \
'gpt-4o(-mini)?(-2024-(08-06|11-20|07-18))?|gpt-4\.1(-mini)?(-2025-04-14)?' .Also search your secrets manager, CI variables, Terraform/Helm values and any admin UI where a model name was typed by hand.
2. Let the API tell you — the shutdown_date field
OpenAI's deprecations page now documents that GET /v1/models and GET /v1/models/{model} return an optional shutdown_date field per model: a YYYY-MM-DD string once a shutdown has been announced, null otherwise. That makes a scheduled check trivial:
curl -s https://api.openai.com/v1/models \
-H "Authorization: Bearer $OPENAI_API_KEY" \
| jq -r '.data[] | select(.shutdown_date != null) | "\(.id)\t\(.shutdown_date)"' \
| sort -k2Run it in CI or a weekly cron, diff against the model IDs you actually use, and fail the job when one of yours gains a date. This catches the next retirement batch, not just this one.
Migration steps
- Inventory every pinned ID with the grep above and the
shutdown_datecheck. Include fine-tunes. - Map each ID to its target. OpenAI's recommendation is
gpt-6-lunafor all four families; if you need more capability than the efficiency tier, evaluategpt-6-solorgpt-6.1-solas well. - Re-run your evaluations against the replacement — quality, latency, cost per task, structured-output validity, and tool-call accuracy — before touching production traffic.
- Roll out behind a canary. Send a slice of traffic to the replacement, compare against stored baselines, then ramp.
- Centralize the model ID so the next forced migration is one edit, not N. This is the fourth OpenAI retirement batch in six months. If model IDs live in dozens of call sites, each batch means touching every one of them. Routing through an OpenAI-compatible gateway turns it into a single config change. As an engineering illustration, byesu is an OpenAI-compatible API gateway where one key reaches current OpenAI models (
gpt-6.1-sol, thegpt-6family includinggpt-6-luna,gpt-5.6) and models from other providers behind the standard OpenAI request shape, so the mapping from "the model we call" to "the model ID upstream" is maintained in the routing layer. A gateway only serves the models it lists — byesu does not offergpt-4oorgpt-4.1, so it is a place to land the migration, not a way to keep calling the retired IDs. Third-party gateways bill at their own published rates; check the provider's own price list rather than assuming parity with OpenAI.
Deadline math
Announced 2026-10-14, shutdown 2027-01-14: that is a three-month notice window, and it spans the December holiday freeze most teams run. Plan the eval and canary phases for November so the cutover is not happening in the last week.
Full upcoming OpenAI retirement calendar (as of 2026-10-26)
All rows are from the OpenAI API Deprecations page. Dates can change; confirm against the official page before you schedule work.
| Shutdown date | Retiring | Official replacement |
|---|---|---|
| 2026-12-11 | gpt-5 / gpt-5-2025-08-07, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3 / o3-2025-04-16, o3-pro | gpt-5.6-sol (gpt-5, o3), gpt-5.6-terra (gpt-5-mini), gpt-5.6-luna (gpt-5-nano), gpt-5.6-sol with reasoning.mode: pro (gpt-5-pro, o3-pro) |
| 2027-01-06 | tts-1, tts-1-hd, gpt-4o-mini-tts (2025-03-20 and 2025-12-15 snapshots) | gpt-realtime-2.1-mini |
| 2027-01-14 | gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-mini (this page) | gpt-6-luna |
| 2027-02-26 | whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize; endpoint /v1/audio/translations | gpt-live-transcribe or gpt-transcribe; v1/realtime/translations |
| 2027-04-01 | gpt-5.3-codex, gpt-5.1 / gpt-5.1-2025-11-13, gpt-5.4-nano / gpt-5.4-nano-2026-03-17 | gpt-6-sol (codex, 5.1), gpt-6-luna (5.4-nano) |
For the batch that already took effect on 2026-07-23 (Codex snapshots, chat-latest aliases, search previews), see Codex / GPT-5 snapshot shutdown.
Source
- OpenAI API Deprecations — https://developers.openai.com/api/docs/deprecations
- OpenAI model pages — https://developers.openai.com/api/docs/models/gpt-4o and https://developers.openai.com/api/docs/models/gpt-6-luna
- OpenAI API pricing — https://developers.openai.com/api/docs/pricing
Always confirm dates, replacement IDs and prices against the official pages above before you migrate; schedules can change.
FAQ
When is GPT-4o being shut down in the OpenAI API?
OpenAI announced on 2026-10-14 that gpt-4o (gpt-4o-2024-08-06, gpt-4o-2024-11-20), gpt-4o-mini (gpt-4o-mini-2024-07-18), gpt-4.1 (gpt-4.1-2025-04-14) and gpt-4.1-mini (gpt-4.1-mini-2025-04-14) will be removed from the API on January 14, 2027. Fine-tuned versions follow the same date.
What replaces gpt-4o and gpt-4o-mini?
OpenAI's recommended replacement for all four families is gpt-6-luna, which it describes as its most efficient model for focused, high-volume tasks. It is a different generation, not a drop-in — re-run your evaluations first.
How can I check programmatically whether a model I use has a shutdown date?
Call GET /v1/models (or GET /v1/models/{model}) and read the optional shutdown_date field: a YYYY-MM-DD string once announced, null otherwise. Alert on it from CI or a scheduled job.
Which other OpenAI models are retiring between December 2026 and April 2027?
2026-12-11: older gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3, o3-pro snapshots. 2027-01-06: tts-1, tts-1-hd, gpt-4o-mini-tts. 2027-01-14: the GPT-4o / GPT-4.1 family. 2027-02-26: whisper-1 and the gpt-4o-transcribe models plus /v1/audio/translations. 2027-04-01: gpt-5.3-codex, gpt-5.1, gpt-5.4-nano. See the calendar above.