SunsettingAI Dev Tools

GPT-4o, GPT-4o mini, GPT-4.1, GPT-4.1 mini

by OpenAI

Announced

2026-10-14

Complete

2027-01-14

Refund window closes

—

Refund status

No refund applicable

On October 14, 2026 OpenAI notified developers that four GPT-4 series models will be removed from the API on January 14, 2027. The deprecations page lists GPT-4o (`gpt-4o-2024-08-06` and `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`), and names one replacement for all four: `gpt-6-luna`. Fine-tuned versions of these models shut down on the same date. Two things stand out. First, the notice is three months, while the same page states a minimum of six months for generally available models unless safety or compliance concerns require a faster timeline, and this entry does not give a reason. Second, the listed price goes down on every row, but `gpt-6-luna` is a reasoning model with no fine-tuning, so the switch is not a plain model string swap.

Refund flow

1
Not applicable. These are pay-per-token API models with no pre-paid balance tied to the model string, so there is no refund category to open.
2
If you are billed for one of these model strings after January 14, 2027, treat it as a billing error and raise it with OpenAI support with the invoice line, rather than as a refund request.
3
Re-budget with the listed prices, per 1M tokens, input then output, as read from OpenAI's model pages on October 14, 2026: gpt-4o $2.50 / $10.00, gpt-4o-mini $0.15 / $0.60, gpt-4.1 $2.00 / $8.00 and gpt-4.1-mini $0.40 / $1.60. gpt-6-luna is $0.10 / $0.50, with cached input at $0.01.
4
On list price that is 25x cheaper input and 20x cheaper output for gpt-4o, 20x and 16x for gpt-4.1, 4x and 3.2x for gpt-4.1-mini, and about 33% and 17% cheaper for gpt-4o-mini. Two caveats before trusting those multiples: gpt-6-luna defaults to medium reasoning effort, so a request can produce reasoning tokens on top of the visible answer, and prompts over 272K input tokens are priced at 2x input and 1.5x output for the full request. Compare a metered day of real traffic, not the rate card.

Migration path

Recommended successor

GPT-6 Luna (gpt-6-luna)

Why this is the closest fit

It is the only replacement OpenAI names on all four rows. Like the models it replaces, it takes text and image input and returns text, and it supports streaming, function calling and structured outputs. OpenAI describes it as its most efficient model for focused, high-volume tasks, which matches how gpt-4o-mini and gpt-4.1-mini were typically used.

What differs from the original

Four things change beyond the model string. First, reasoning: gpt-4o and gpt-4.1 had no reasoning step, while gpt-6-luna supports reasoning.effort values none, low, medium, high, xhigh and max, with medium as the default. Set effort explicitly if you want the old no-reasoning behaviour. Second, function calling on Chat Completions: the gpt-6-luna page says Chat Completions supports function calling only with reasoning_effort set to none, so a Chat Completions app that calls tools has to set that or move to the Responses API. Third, fine-tuning: all four old model pages list fine-tuning as supported, and the gpt-6-luna page lists it as not supported. A fine-tuned gpt-4o-mini or gpt-4.1-mini goes away on January 14, 2027 and the named replacement cannot be tuned, so the tuned behaviour has to move into the prompt. Fourth, predicted outputs are listed as supported on all four old model pages and are not listed on the gpt-6-luna page. Context grows: 1,050,000 tokens and 128,000 max output, against 128,000 and 16,384 on gpt-4o and gpt-4o-mini, and 1,047,576 and 32,768 on gpt-4.1 and gpt-4.1-mini.

Alternatives by use case

You use gpt-4o or gpt-4o-mini to expand or rewrite prompts before sending them to a video model

Run the same prompt set on gpt-6-luna with reasoning effort set to none and to low, and compare both the output and the token count, before January 14, 2027

You run a fine-tuned gpt-4o-mini or gpt-4.1-mini

There is no tunable replacement on the deprecation row. Collect the examples you trained on and test them as few-shot examples in a gpt-6-luna prompt now, while the tuned model still works for side by side comparison

You call tools through Chat Completions

Set reasoning_effort to none on gpt-6-luna, or plan a move to the Responses API, because the model page says Chat Completions function calling requires it

You use gpt-4o to describe frames or thumbnails from your clips

gpt-6-luna also accepts image input, so the change is the model string plus an effort setting. Re-check any stored expected outputs, since the old snapshot pins are what is being removed

What this tool meant

GPT-4o and GPT-4.1 were the default models behind a large share of 2024 and 2025 AI apps, and gpt-4o-mini was the low-cost workhorse for captioning, tagging and prompt rewriting. They also get three months of notice, against the six-month minimum OpenAI writes down for generally available models. Unlike the October 23, 2026 batch, where the small-model rows got more expensive, every row here gets cheaper on list price. The cost of this migration is in behaviour, not price: a non-reasoning model with fine-tuning is being replaced by a reasoning model without it.