OpenAI sets January 14 deadline for GPT-4o and GPT-4.1 API migrations
Teams have three months to move production workloads, including fine-tuned versions, and the recommended replacement is not merely a new model name.
The deadline
On October 14, OpenAI scheduled five GPT-4-series snapshots for removal from its direct API: two GPT-4o versions, GPT-4o mini, GPT-4.1 and GPT-4.1 mini. Access ends on January 14, 2027, and fine-tuned versions follow the same schedule.
The contract changes
OpenAI recommends GPT-6 Luna for all five models. Its standard rate is $0.10 per million input tokens and $0.50 per million output tokens, but Luna introduces reasoning settings and different function-calling behavior across Chat Completions and the Responses API. A slug change is not a migration test.
What we would do
At Virtual Arc, we would inventory model IDs, fine-tuned bases and provider paths this week, then add centralized routing and shadow representative traffic. Check deadlines by provider: Microsoft Foundry currently publishes April 14, 2027 for the corresponding versions, rather than the direct OpenAI API date.
Our view is that this matters more than another flashy model launch because it turns architecture debt into an outage with a date. A team that embedded model IDs, prompts and tool behavior directly into business logic does not have a stable integration; it now has a three-month migration project. At Virtual Arc, we would not accept GPT-6 Luna as a drop-in replacement simply because the vendor recommends it. We would first introduce centralized model routing and inventory every direct call, fine-tuned base model and provider path; then shadow representative production traffic and measure completed-task success, p95 latency, output-token volume, tool-call validity and structured-output failures. For routine, latency-sensitive routes, we would start Luna without additional reasoning and enable higher effort only where evaluations improve cost per completed task. This migration is worth starting now, but production traffic should not move until the replacement clears the same operational gates as the existing model.