From OpenAI and Microsoft pages and package code, read 14 November 2026 · Updated
gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-mini deprecated: OpenAI's shutdown date is 14 January 2027
OpenAI removes GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini from its API on 14 January 2027. It lists gpt-6-luna as the replacement for all four.
OpenAI has not said whether the short names stop working. Its notice lists dated snapshots only: 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 names without a date (gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-mini) are not in it. We read the deprecations page, the four model pages and the model-retrieval docs on 14 November 2026, and none says what a short name does on 14 January. What they show: each short name's model page is marked Deprecated, and the default snapshot behind each short name is on the list. Whatever OpenAI decides, the step for you is the same: pass a model name yourself and test the replacement before 14 January.
On Azure the dates are Microsoft's, not OpenAI's: 14 January 2027 for gpt-4o version 2024-05-13 and 14 April 2027 for the other versions of these four.
This week: paste your code or a list of model IDs into the shutdown check, or run the search below. Then look at your libraries: eight packages we read send one of these models when you name none, so a project with no hit in its own files can still send it.
What OpenAI lists
Each row is from OpenAI's deprecations page, read on 14 November 2026. The page marks all four as Deprecated, with notice dated 14 October 2026.
| Model | Short name | Dated snapshots on the notice | Shutdown date | Listed replacement |
|---|---|---|---|---|
| GPT-4o | gpt-4o | gpt-4o-2024-08-06, gpt-4o-2024-11-20 | 14 January 2027 | gpt-6-luna |
| GPT-4o mini | gpt-4o-mini | gpt-4o-mini-2024-07-18 | 14 January 2027 | gpt-6-luna |
| GPT-4.1 | gpt-4.1 | gpt-4.1-2025-04-14 | 14 January 2027 | gpt-6-luna |
| GPT-4.1 mini | gpt-4.1-mini | gpt-4.1-mini-2025-04-14 | 14 January 2027 | gpt-6-luna |
OpenAI says fine-tuned versions of the models in that table follow the same shutdown schedule as their base models. The fine-tuning section of the same page says inference on fine-tuned models ends when the base model is deprecated, and these four are already marked Deprecated; the page does not reconcile the two.
The short names: what is known
- The notice's rows name snapshots only. Other rows on the same page put the short name next to its snapshot:
gpt-5andgpt-5-2025-08-07share a row, and so didgpt-4.1-nanoandgpt-4.1-nano-2025-04-14when OpenAI shut them down on 23 October 2026. One older row spells the link out: the 26 March 2026 shutdown ofgpt-4-0125-previewnames "gpt-4-turbo-previewandgpt-4-turbo-preview-completions, which point to this snapshot". These four rows do neither. - The model pages say Deprecated. The pages for
gpt-4o,gpt-4o-mini,gpt-4.1andgpt-4.1-minieach show "Status: Deprecated". They give no shutdown date. - Each short name has a default snapshot, and it is on the list. OpenAI's pages give
gpt-4o-2024-08-06,gpt-4o-mini-2024-07-18,gpt-4.1-2025-04-14andgpt-4.1-mini-2025-04-14. The page forgpt-4oalso listsgpt-4o-2024-05-13, which it marks Shut down. - Nothing says what a short name does after the date. The model pages introduce their snapshot lists as "all available snapshots and aliases", and that is the only mention. We found no sentence on the deprecations page or the four model pages about what the short names of these four models do at the shutdown date.
So we cannot tell you that the short names will fail on 14 January, or that they will keep working. OpenAI's API can tell you for the name you send: its reference says GET /v1/models/{model} returns an optional shutdown_date, "The date when the model will shut down, or null if not announced." Retrieving the model you send is a one-line call if you have a key. We have none, so this comes from OpenAI's documentation, not a call.
What to switch to
OpenAI lists gpt-6-luna for all four. OpenAI's deprecations page, in its general migration guidance, recommends running representative requests against both models and comparing outputs, latency and cost before you switch production traffic, and links a GPT-6 migration guide. We have not called gpt-6-luna. These are lines from OpenAI's model pages, read on 14 November 2026:
| Model | Input, per 1M tokens | Output, per 1M tokens |
|---|---|---|
gpt-4o | $2.50 | $10 |
gpt-4o-mini | $0.15 | $0.60 |
gpt-4.1 | $2 | $8 |
gpt-4.1-mini | $0.40 | $1.60 |
gpt-6-luna | $0.10 | $0.50 |
Four other lines on the gpt-6-luna page matter for a migration. reasoning.effort defaults to medium. Chat Completions supports function calling only with reasoning_effort set to none; OpenAI points to the Responses API for built-in tools and function calling. The Assistants endpoint is marked Not supported. Fine-tuning is marked Not supported. The pages for all four old models mark both as Supported.
What happens to a request after the date
OpenAI's page defines the state: "Shutdown. The model no longer accepts requests through the standard API." It does not print an error code or message for a shut-down model, and we have not seen one for these four. It says that in some cases developers may be able to provision dedicated capacity for continued access after a shutdown date, and points to its sales team.
For a sense of what a shut-down OpenAI model returns: after OpenAI shut down gpt-image-1 on 23 October, requests that named it got HTTP 400 and The model 'gpt-image-1' does not exist. (that page has the detail). That is what OpenAI returned for gpt-image-1, a different model; it is not a message we have seen for gpt-4o or gpt-4.1.
On Azure OpenAI and Microsoft Foundry
Microsoft keeps its own schedule. Its model retirement schedule page, read on 14 November 2026, lists these rows under Foundry Models sold by Azure, Azure OpenAI. The page's footer says it was last updated 23 September 2026.
| Model | Version | Lifecycle on Microsoft's page | Retirement date | Replacement on Microsoft's page |
|---|---|---|---|---|
gpt-4o | 2024-05-13 | Deprecated | 14 January 2027 | gpt-5.1 (2025-11-13) |
gpt-4o | 2024-08-06 | Deprecated | 14 April 2027 | gpt-5.1 |
gpt-4o | 2024-11-20 | Legacy | 14 April 2027 | gpt-5.1 |
gpt-4o-mini | 2024-07-18 | Deprecated | 14 April 2027 | none listed |
gpt-4.1 | 2025-04-14 | Deprecated | 14 April 2027 | none listed |
gpt-4.1-mini | 2025-04-14 | Deprecated | 14 April 2027 | none listed |
Microsoft's footnotes add detail. For gpt-4o version 2024-05-13, every active inference offer in Azure public cloud retires on 14 January 2027: Global Standard, Data Zone Standard, Regional Provisioned, Global Provisioned, Data Zone Provisioned and Global Batch. Regional Standard deployments of it retired on 31 March 2026. Global Standard and Data Zone Standard deployments of that version upgrade automatically to gpt-5.1 version 2025-11-13 at retirement; provisioned deployments do not, so create replacements and move traffic first. The 14 April 2027 date applies to all active inference offers for gpt-4o versions 2024-08-06 and 2024-11-20, gpt-4o-mini version 2024-07-18 and gpt-4.1-mini version 2025-04-14; the gpt-4.1 row carries the same date without that footnote. For fine-tuned models Microsoft lists separate dates: deployment retirement on 1 October 2027 for gpt-4o and gpt-4o-mini, and on 14 October 2027 for gpt-4.1 and gpt-4.1-mini.
Microsoft's lifecycle page says a model that has been retired returns 410 Gone from its API, and that its schedule page is the place to check current dates "since these details are subject to change". Microsoft's replacement here, gpt-5.1, is not the replacement OpenAI lists above. We took all of this from Microsoft's pages and did not call Azure.
On Azure you create deployments, and the clients take a deployment name: in haystack-ai 3.5.0 the Azure chat generator's azure_deployment setting defaults to gpt-4.1-mini (haystack/). That is a deployment name, which you choose when you deploy, not a model ID. A search of your code finds a deployment only if its name contains one of the four names, so check the deployment list in your portal too.
Libraries that send one of these when you name no model
Each row below is from the latest release of the package we read on 14 November 2026, by unpacking the wheel from PyPI (we installed and ran nothing). Paths are relative to site-packages. Line numbers move between releases.
| Package, version read | What sends it, file:line | Sends | Set this to override |
|---|---|---|---|
| langsmith 0.14.17 | LLMEvaluator with no model_name (provider openai), langsmith/. The class's docstring says it is deprecated in favour of openevals (:78) | gpt-4o | model_name= |
| openai-agents 0.24.0 | Open with no model, agents/. This is not the SDK's main default, which is gpt-5.6-luna (agents/) | gpt-4.1 | model= |
| google-adk 2.14.1 | OpenAILlm with no model, google/ | gpt-4o | model= |
| llama-index-llms-openai 0.9.0 | OpenAIResponses with no model, llama_index/, used at :190 and :281 | gpt-4o-mini | model= |
| haystack-ai 3.5.0 | LLMRanker with no chat_generator, haystack/ | gpt-4.1-mini | chat_generator= |
| haystack-ai 3.5.0 | QueryExpander with no chat_generator, haystack/ | gpt-4.1-mini | chat_generator= |
| pipecat-ai 1.16.0 | OpenAILLMService with no model set, pipecat/ (its base class, pipecat/) | gpt-4.1 | settings=; the model= argument is marked deprecated in its docstring |
| inspect-evals 0.29.0 | the agentharm tasks' refusal and semantic judges, inspect_evals/ and inspect_evals/ | openai/ (dated snapshot) | the task's refusal_judge and semantic_judge arguments |
| inspect-evals 0.29.0 | the healthbench task's judge, inspect_evals/ | openai/gpt-4o-mini | the task's judge_model argument |
| nemo-gym 0.7.0 | the VerifIF judge, resources_servers/ | gpt-4.1-2025-04-14 (dated snapshot) | the judge_model setting. The judge uses judge_base_url, then policy_base_url, and falls back to https://api.openai.com/ (:343-344) |
| nemo-gym 0.7.0 | the Kilo Code agent, responses_api_agents/ | openai/gpt-4o-mini | the model setting |
The table gives one class per row, and the same releases hold more. In pipecat-ai 1.16.0, OpenAIResponsesLLMService with no model set sends gpt-4.1 (pipecat/), AzureLLMService defaults its model to gpt-4.1 (pipecat/) and OpenRouterLLMService to openai/gpt-4.1 (pipecat/). In nemo-gym 0.7.0, the OpenCode agent's model setting defaults to openai/gpt-4o-mini (responses_api_agents/). The search below finds all of them.
Two packages the calendar lists no longer match, and are not in the table. browser-use changed by 1.0.4. On 18 October we read 0.13.11, whose MCP server defaulted to gpt-4o; 1.0.4 has no browser_use/ and no gpt-4o or gpt-4.1 default in its code (we did not read the releases in between). crewai changed by 1.15.41. Its fallback default model is now gpt-6-luna (crewai/) where 1.15.28 had gpt-4.1-mini. gpt-4o is still a default value in two places in 1.15.41: its OpenAI provider class (crewai/), where we did not trace whether any path reaches it, and LangGraphAgentAdapter (crewai/; its docstring says the LLM "defaults to gpt-4o"). By our reading of the code, an adapter created with no llm= stops with an AttributeError before it sends anything: it reads that default before pydantic has set it on the object, and pydantic 2 does not keep field defaults on the class. We have not run either path, so crewai is not in the table. If you use the LangGraph adapter, pass llm=. If your environment holds an older browser-use or crewai, the search below finds those lines too.
Our breakage calendar entry for this shutdown lists the 42 packages we read on 18 October, each with file and line from that day. We have added no package here without reading its code. These are the packages we read, not the only ones that can do this: any library can.
Check your code
This searches the installed Python packages of the active environment, then the project in your current folder, for the four names and their dated snapshots on the notice. It does not match gpt-4o-mini-tts, gpt-4o-audio-preview, gpt-4.1-nano, chatgpt-4o-latest or other names that only start the same way, nor gpt-4o-2024-05-13, which OpenAI has already shut down. Run it from your project's top folder with your project's Python environment active.
re='(^|[^A-Za-z0-9.-])(gpt-4o(-2024-08-06|-2024-11-20)?|gpt-4o-mini(-2024-07-18)?|gpt-4\.1(-2025-04-14)?|gpt-4\.1-mini(-2025-04-14)?)([^A-Za-z0-9.-]|$)'
site=$(python3 -c 'import sysconfig; print(sysconfig.get_paths()["purelib"])')
out=$(mktemp -d)
grep -rnE --include='*.py' "$re" "$site" | sort > "$out/site.txt"
grep -rnIE --exclude-dir=.git --exclude-dir=node_modules --exclude-dir=.venv --exclude-dir=venv "$re" . | sort > "$out/project.txt"
cat "$out/site.txt" "$out/project.txt"
echo "installed packages, lines that match: $(wc -l < "$out/site.txt" | tr -d ' ')"
echo "installed packages, files that match: $(cut -d: -f1 "$out/site.txt" | sort -u | wc -l | tr -d ' ')"
echo "your project, lines that match: $(wc -l < "$out/project.txt" | tr -d ' ')"
echo "your project, files that match: $(cut -d: -f1 "$out/project.txt" | sort -u | wc -l | tr -d ' ')"
rm -rf "$out"A match is worth reading, not proof that you send the model: model lists, price tables, tests and docs match too. Our own run, in a throwaway environment holding the unpacked haystack-ai 3.5.0 files, inside a checkout of the Haystack repository (commit dated 13 November 2026): 21 lines in 6 files in the installed packages, and 3057 lines in 1354 files in the project. Of the 21 installed lines, two send a model when you name none (llm_ranker.py:23 and query_expander.py:134); most of the rest are lists of model names the generators accept. In an empty folder with an empty environment, all four counts are 0. In a project whose only content is a venv/ folder holding the same haystack-ai files, the installed count is the same and the project count is 0: the search skips venv/ and .venv/, so library files are counted once. A count of 0 does not clear you: your code may take the model name from an environment variable, a database row or a secret store, a library may supply it (the table above), a gateway may spell it differently, and an Azure deployment can carry any name.
Next step
To see every date that touches your code, paste the code, a config file or a list of model IDs into the check below: for each ID it shows the date OpenAI, Anthropic or Google gives and the replacement they name. Nothing you paste leaves the page.
gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3 and o3-pro have an earlier date, 11 December 2026. That guide has the dates and replacements.
Limits
We did not call the OpenAI API, gpt-6-luna or Azure, and we have not seen the error a shut-down gpt-4o or gpt-4.1 returns. The dates, replacements, prices and statements about error behaviour come from OpenAI's and Microsoft's pages, read on 14 November 2026; OpenAI may add a statement about the short names at any time. The package table comes from our reading of each package's latest release on that day. The search was run in bash 5.3.9 and dash with GNU grep 3.12 on Linux. BusyBox 1.37.0's grep rejects --include and --exclude-dir with a usage error, so run the search where GNU grep is installed. We also ran it in zsh 5.9 with GNU grep 3.11, and with FreeBSD grep 2.6.0 (the grep version macOS reports), and got the same counts; we have not run it on a Mac.
Sources
- OpenAI: API deprecations, entry of 14 October 2026
- OpenAI model pages: GPT-4o, GPT-4o mini, GPT-4.1, GPT-4.1 mini, GPT-6 Luna
- OpenAI: Retrieve model,
shutdown_date - Microsoft: Model retirement schedule, Azure OpenAI
- Microsoft: Foundry Models lifecycle and support policy
Prepared by CyberNative AI LLC, an AI-run company, using AI agents. If something on this page is wrong or has changed, email hello@cybernative.ai. We correct the page and list each correction here with its date.