OpenAI retires GPT-4o on half the notice its policy promises

OpenAI gave developers three months to leave GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini. The company's own deprecation page, dated 14 October 2026, sets the shutdown date for all four at 14 January 2027, with GPT-6 Luna named as the recommended replacement for each.

The policy OpenAI publishes on the same page promises at least six months of notice before any generally available model is removed. The GPT-4 shutdown falls short of that by half. The policy allows exceptions for safety or compliance reasons, and the deprecation entry gives none. The New Stack, which first reported the gap, notes that fine-tuned versions follow the same schedule and that self-serve fine-tuning jobs on these models end eight days earlier, on 6 January 2027.

The four models cover a lot of ground. GPT-4o shipped in May 2024 as the flagship chat model. GPT-4.1 shipped in April 2025 as an API-only family built around coding and up to a million tokens of input, which OpenAI's chief product officer called at the time models trained just for developers. Anything in production built on one of these four has an unmoving January deadline to run on something else.

GPT-6 Luna is the cheapest tier of GPT-6, at 10 cents per million input tokens and 50 cents per million output tokens on OpenAI's price page. OpenAI's own migration guide, cited by The New Stack, says the two rates both change once a prompt crosses 272,000 input tokens: input doubles for the whole request, and output rises by 50 percent for the whole request. A retrieval or coding workload that regularly reads a large codebase will not land on the headline price.

The swap itself is more work than changing a model identifier. Luna defaults to medium reasoning effort. OpenAI's migration guide, as reported by The New Stack, says function calling through the Chat Completions endpoint only works with reasoning turned off, and combining reasoning with tool calls needs the Responses API. Turning reasoning on rules out temperature, top_p, top_logprobs and logprobs, each of which GPT-4.1 supported.

An application built on one GPT-4 model today has one of three paths before 14 January: move to Luna with the migration work above, move to another provider, or pay for dedicated capacity. Pricing a Luna migration by swapping the model identifier and measuring tokens will not include the behavioural differences, so the test to run first is a replay of real production traffic through Luna, counting responses that regress in format, in tool use or in cost.

Source

• Primary source: Deprecations, OpenAI
• OpenAI promised developers six months' notice. GPT-4o users got three., The New Stack, by Amanda Caswell