OpenAI has set a January 14, 2027 shutdown for five GPT-4 API versions. Its October 14, 2026 notice also covers fine-tuned versions built on them. Businesses whose software uses these models will need to move to a supported replacement.
The first check is which model an application actually uses. For a purchased chatbot or workflow, ask the supplier for its model dependencies and migration plan; a “GPT-4” label alone is not specific enough.
Which versions are affected?
The retirement notice names these dated releases, also called snapshots:
GPT-4o:
gpt-4o-2024-08-06andgpt-4o-2024-11-20.GPT-4o mini:
gpt-4o-mini-2024-07-18.GPT-4.1:
gpt-4.1-2025-04-14.GPT-4.1 mini:
gpt-4.1-mini-2025-04-14.
OpenAI recommends GPT-6 Luna for all five. This is not a blanket January deadline for every GPT-4 model: other versions have their own dates. The notice specifies neither a shutdown hour nor an automatic switch to the replacement.
Why a simple swap may not work
Custom-trained models need a different plan. Fine-tuning means training a model on examples for a particular task. Luna’s model documentation says it does not support fine-tuning. A business cannot therefore assume it can recreate its existing customized model by training Luna on the same examples.
The practical implication is to evaluate whether instructions and examples supplied with each request can reproduce the behaviour the business needs. If they cannot, the team needs another approach before the deadline. A recommended replacement is not evidence that learned customization carries over.
Applications that take actions need a compatibility check. Function calling lets a model ask software to perform an action. OpenAI’s migration guide says Luna supports this through Chat Completions only when reasoning_effort is set to "none". To combine reasoning with tools, the documented route is the Responses API. Luna’s default reasoning setting is medium, so the reasoning setting needs an explicit check.
Teams considering Luna can evaluate the existing Chat Completions integration with reasoning disabled, or assess a move to Responses for reasoning with tools. Fixing request compatibility and checking the quality of the resulting actions are separate jobs.
What to check before switching
Inventory the model versions behind both everyday features and infrequently run scheduled jobs, including any fine-tunes. Confirm the hosting provider’s own retirement notice rather than applying this OpenAI API date to every service selling similarly named models.
Then compare the proposed replacement on representative tasks, including the difficult examples that motivated custom training. Following OpenAI’s task-specific evaluation guidance, set acceptance criteria for correct answers, required output formats and tool actions. Also check response time and billed usage, and choose a supported fallback before switching production traffic.
Reporting note: Prepared with AI assistance from official OpenAI documentation checked on October 15, 2026. The migration implications are analysis; no hands-on model testing was performed.