GPT-4o and GPT-4.1 retire January 14, GPT-4.1 nano in a week: what moving to GPT-6 Luna costs

In short

• gpt-4.1-nano and the 2024-05-13 GPT-4o snapshot stop on October 23. Everything else in 4o and 4.1 stops January 14.
• GPT-6 Luna takes a $700 GPT-4o month to $30, but only with reasoning set to none. Its default is medium.
• Nano has no cheaper OpenAI successor that isn't also being retired. Luna is the closest, at $30 against $28.

Two notices, two dates

The deprecations page now has two entries that between them cover every GPT-4o and GPT-4.1 text model still running. They don't share a date or a replacement.

OpenAI deprecations page, read October 16, 2026. The October 23 rows were announced April 22; the January 14 rows on October 14. Fine-tuned models follow their base model's date.

The October 14 entry lists dated snapshots only. It never names the bare gpt-4o, gpt-4o-mini, gpt-4.1 or gpt-4.1-mini aliases. The GPT-4o model page lists 2024-08-06 as the default snapshot, so we'd assume the alias stops working with it. OpenAI hasn't said so in writing, and we wouldn't bet a production service on the alias quietly being repointed.

The nano entry does name its alias, so plain gpt-4.1-nano calls fail next Friday. Realtime and audio 4o models are on a third notice, ending January 20, and chatgpt-4o-latest has been gone since February.

What the same month costs on each

We priced a 10:1 input-heavy month, 200M tokens in and 20M out, the shape of most classification, extraction and support traffic. No caching, no batch, and no reasoning tokens. That last condition matters more than the rest, and we come back to it below.

Our arithmetic from OpenAI's pricing page, read October 16, 2026. Standard tier. GPT-6 Luna, GPT-6.1 Sol and GPT-5.6 Luna bill the whole request at 2x input and cache and 1.5x output above 272K input tokens; GPT-4o and 4.1 have no such step.

For the two big models this is the cheapest forced migration we've covered. A $700 GPT-4o month becomes $30 on GPT-6 Luna, a little over 4 cents on the dollar. GPT-4.1 drops from $560, and 4.1 mini from $112. Even at the 2x long-context rate, a 400K-token GPT-4.1 prompt costs about a tenth as much on Luna.

GPT-4o mini is where it gets close. $42 to $30 is a 29% saving, and it's the saving that disappears first.

Luna thinks unless you tell it not to

GPT-4o never produced reasoning tokens. GPT-6 Luna is a reasoning model with reasoning_effort set to medium by default, and OpenAI's reasoning guide bills those tokens as output at $0.50 per million. Change only the model string and you get medium.

We can't tell you how many reasoning tokens your prompts will produce. That depends on the task. What we can tell you is where each saving runs out, on the month above:

• From GPT-4o mini: after 24M reasoning tokens a month, 1.2x the visible output. Medium effort on anything harder than a lookup clears that easily.
• From GPT-4.1 mini: after 164M, 8.2x the visible output.
• From GPT-4o: after 1,340M, 67x. You won't get there.

Luna accepts none, which is the setting that matches what you had. There's a second reason to set it: the Luna model page says function calling through Chat Completions only works with effort at none. Plenty of 4o mini code is exactly that, Chat Completions with tools, and it will break on a straight swap rather than just cost more.

Caching changes too. 4o mini read cached input at $0.075 and charged nothing to write it. Luna reads at $0.01 but bills a $0.125 cache write, a quarter more than plain input. Prefixes you reuse a lot come out well ahead. Prefixes you write once and barely reuse can cost a little more than before.

GPT-4.1 nano: the one migration that costs more

Nano is the only model in the batch with no cheaper OpenAI successor that isn't itself on a shutdown list. At $0.10 in and $0.40 out it ran our month for $28. OpenAI points it at GPT-5.6 Luna. That's $64, more than double the bill.

GPT-6 Luna is the better target even though the notice doesn't name it. Same $0.10 input, $0.50 output instead of $0.40, so $30 on our month, 7% more than nano with reasoning off. GPT-5 nano is cheaper still, $18 on our month at $0.05 in and $0.40 out, but it shuts down on December 11, so it buys you seven weeks. Of the OpenAI models with no shutdown date, Luna is the cheapest.

The same April entry retires the original gpt-4o-2024-05-13 snapshot, the $5 / $15 one, and a set of old fine-tunes (ft-gpt-4, ft-gpt-3.5-turbo, ft-o4-mini and others). If anything in your codebase still pins that date string, it goes on October 23 too.

Fine-tunes and Azure

Fine-tuned 4o and 4.1 models stop on their base model's date, and you can't move the fine-tune across: GPT-6 Luna doesn't support fine-tuning. OpenAI also stops active existing customers from creating new fine-tuning jobs on any model on January 6, eight days before the shutdown. Organisations with no fine-tuned inference in the previous 60 days lost that on July 2. A fine-tuned 4o mini doing a narrow task has to become a prompted Luna call. Budget time for evals, because the fine-tune's gains won't carry over on their own.

Azure is on a slower clock. Microsoft's retirement schedule lists April 14, 2027 for the 2024-08-06 and 2024-11-20 GPT-4o versions, 4o mini and all three GPT-4.1 models, and December 9 for 2024-05-13. Fine-tuned deployments run until October 2027. That page was last updated September 23, before OpenAI's notice, so the dates could still move. Azure retirements can't be extended, and a retired deployment returns 410 Gone.

Our order of work

1. Grep for gpt-4.1-nano and 2024-05-13 today. Those have a week. Move them to gpt-6-luna with effort none, not the GPT-5.6 models the notice names.
2. Move GPT-4o and GPT-4.1 traffic to Luna at none first. Most of it will hold, and it's 19x to 23x cheaper before you change anything else. Raise effort only on the prompts that fail, and only as far as low until you've measured.
3. Watch 4o mini traffic closely. It has the thinnest margin, and one default left at medium is enough to make the move cost more.
4. Send whatever Luna can't handle to GPT-6.1 Sol. At $600 on our month it's still under GPT-4o, but it has no none effort and no tool calling in Chat Completions, so that path means moving to the Responses API.

We went through Luna's card in more detail in our GPT-6 Luna pricing post. To price your own mix, put your token counts into the calculator and add the reasoning tokens to the output side.

Sources

• OpenAI: deprecations - October 14 and April 22 entries, realtime and audio entry, fine-tuning cutoffs
• OpenAI: API pricing - every rate in the cost table, long-context rule, cache writes
• OpenAI: GPT-6 Luna model page - default effort, function calling limit, no fine-tuning
• OpenAI: reasoning guide - reasoning tokens billed as output
• Microsoft: Azure OpenAI model retirement schedule and retirement policy