OpenAI Is Switching Off GPT-4o, GPT-4.1 and GPT-5-nano in Its API. What Breaks When, What It Costs, and How to Migrate

By Thorsten Meyer | November 1, 2026 | ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Somewhere in your company there is probably a line like model: "gpt-4o". It sits in a config file, a test fixture or a no-code automation that someone built in 2024 and nobody has touched since. On 14 January 2027, OpenAI switches that model off in its API. Five weeks earlier, on 11 December, GPT-5, GPT-5 mini and GPT-5-nano go. And some older IDs already stopped working on 23 October, which means a few requests may be failing right now.

In the same week, OpenAI rewrote the Model Spec, its public rulebook for how its models should behave. It drops most of the detailed style guidance and tells the model to read instructions for their intent. Two changes, one practical question: what do you have to do before your app starts answering differently, or not at all? This guide sorts out the dates, the price differences, the behavior traps, and a test plan you can run before the deadline.

Key takeaways

• Three dates matter most. 11 December 2026: gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3 and o3-pro leave the API. 14 January 2027: the remaining GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini snapshots follow. Already gone since 23 October: among others gpt-4.1-nano, the first GPT-4o snapshot, gpt-4-turbo, gpt-3.5-turbo, o1 and o4-mini.
• Fine-tuned models die with their base. OpenAI says fine-tuned versions follow the base model’s shutdown schedule. The replacements it lists do not support fine-tuning.
• ChatGPT users already lost GPT-4o in February. OpenAI removed GPT-4o, GPT-4.1 and GPT-4.1 mini from ChatGPT on 13 February 2026. This round is about the API, which means about the apps and bots built on it.
• Cheaper per token in January, pricier in December. The listed replacement for the GPT-4o family, gpt-6-luna, costs $0.10 per million input tokens against $2.50 for GPT-4o. But gpt-5-nano users move to gpt-5.6-luna at four times the input price, and gpt-5-mini users to gpt-5.6-terra at eight times.
• The replacements think before they answer. They are reasoning models. Hidden reasoning tokens are billed as output, some sampling parameters must go, and a tight output limit can end in an empty, billed answer.
• The new Model Spec stops promising a house style. The 29 October version removes most style sections, including the default to format answers in Markdown. If your app depends on a format, say so in the request. OpenAI itself writes that its models “do not yet fully reflect” the Spec.

Products Worth Considering

Samsung 32-inch Odyssey OLED G8 (G80SH) 4K QD-OLED Gaming Monitor

AmazonView Latest Price

Samsung 27-inch Odyssey QD-OLED G8 4K 240Hz Gaming Monitor

AmazonView Latest Price

LG UltraGear 32GX850A-B 32-inch 4K UHD Glossy OLED Gaming Monitor

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

The shutdown calendar

All dates below come from OpenAI’s deprecations page, which I checked on 1 November 2026. OpenAI says it emails affected customers and documents each deprecation there.

• Already shut down (23 October 2026). Announced in April: gpt-3.5-turbo, gpt-4, gpt-4-turbo, gpt-4.1-nano, the original GPT-4o snapshot gpt-4o-2024-05-13, o1, o1-pro, o3-mini, o4-mini and the image model gpt-image-1. If any of these still sit in your code, those calls fail today.
• 30 November 2026. Not a model, but it matters for testing: OpenAI shuts down its Evals platform, the Agent Builder and the reusable prompt objects (v1/prompts). Existing evals became read-only on 31 October.
• 11 December 2026. Notified on 11 June: gpt-5 goes to gpt-5.6-sol, gpt-5-mini to gpt-5.6-terra, gpt-5-nano to gpt-5.6-luna. o3 moves to gpt-5.6-sol; gpt-5-pro and o3-pro to gpt-5.6-sol with its pro reasoning mode.
• 14 January 2027. Notified on 14 October: gpt-4o-2024-08-06, gpt-4o-2024-11-20, gpt-4o-mini-2024-07-18, gpt-4.1-2025-04-14 and gpt-4.1-mini-2025-04-14. The recommended replacement for all of them is gpt-6-luna.
• Later in 2027. Text-to-speech models such as tts-1 on 6 January, older audio and realtime models including the GPT-4o audio variants on 20 January, whisper-1 and the GPT-4o transcription models on 26 February, and gpt-5.1, gpt-5.4-nano and gpt-5.3-codex on 1 April.

Two details are easy to miss. First, the tables list dated snapshots, not the short names most people type. The model pages for the plain gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini names already show the status “Deprecated”, and the snapshots behind them are on the January list. Treat the short names as part of that date. Second, the notice period. The same page promises at least six months’ notice for generally available models, unless safety or compliance concerns require a faster timeline. From 14 October to 14 January is three months. The entry does not say which rule applies. If you have a contract or a committed spend with OpenAI, that is worth a question to your account team. OpenAI also says some customers may be able to buy dedicated capacity for continued access after a shutdown.

What the replacements cost

The official pricing page lists standard prices per million tokens. The comparison below uses the short-context prices, which apply to prompts up to 272,000 tokens. I checked it on 1 November 2026; prices change, so check again before you budget.

• GPT-4o → GPT-6 Luna: $2.50 input and $10.00 output for GPT-4o, against $0.10 and $0.50. Cached input falls from $1.25 to $0.01.
• GPT-4o mini → GPT-6 Luna: $0.15 and $0.60, against $0.10 and $0.50.
• GPT-4.1 → GPT-6 Luna: $2.00 and $8.00, against $0.10 and $0.50.
• GPT-4.1 mini → GPT-6 Luna: $0.40 and $1.60, against $0.10 and $0.50.
• GPT-5-nano → GPT-5.6 Luna: $0.05 and $0.40, against $0.20 and $1.20. That is four times the input price and three times the output price.
• GPT-5 mini → GPT-5.6 Terra: $0.25 and $2.00, against $2.00 and $12.00. Eight times the input price, six times the output price.
• GPT-5 → GPT-5.6 Sol: $1.25 and $10.00, against $4.00 and $20.00. The page notes that GPT-5.6 Sol’s promotional pricing runs at least through 21 November 2026.

So the January move looks like a discount, and the December move looks like a price increase. Both readings are only half true, because the price per token is not the price per task. GPT-4o and GPT-4.1 answer directly. All the listed replacements are reasoning models, and OpenAI’s reasoning guide says their hidden reasoning tokens “are billed as output tokens”. A model with a lower price per token can still cost more per answer if it thinks a lot. The opposite also happens: in my own newsroom test of a sibling model, GPT-6.1 Terra was cheap per task, and its maximum setting was the worst.

A second point is my own reading of the price list, not an OpenAI recommendation. For the December group, OpenAI lists GPT-5.6 models as replacements, while the newer GPT-6 family is cheaper per token on the same page: gpt-6-luna at $0.10 and $0.50, gpt-6.1-terra at $1.00 and $5.00, gpt-6.1-sol at $2.00 and $10.00. If you migrate anyway, test a GPT-6 option alongside the listed one. Regional processing endpoints add 10% for models released since March 2026. For the wider price fight between the labs, see the AI price war and its fine print.

What changes besides the price

OpenAI’s migration guide for the GPT-6 family lists the technical changes. These are the ones most likely to break an app that was built for GPT-4o.

• Reasoning effort is a new dial. gpt-6-luna supports the efforts none, low, medium (the default), high, xhigh and max. GPT-4o never reasoned, so the default is not a like-for-like swap. My reading: start with none or low for simple, high-volume tasks, then compare.
• Some parameters must go. When reasoning effort is not none, the guide says to remove temperature, top_p and top_logprobs. Old code that sets a temperature can fail or need changes.
• Tool calls depend on the API you use. GPT-6 Luna supports function calling in Chat Completions only with reasoning effort none. For reasoning with tools, OpenAI points to its Responses API.
• Tight output limits can bite. If the model hits max_output_tokens while it is still reasoning, the response comes back “incomplete”, possibly before any visible text. OpenAI warns that you can pay for input and reasoning tokens without getting an answer, and suggests reserving at least 25,000 tokens for reasoning and output when you start experimenting.
• No fine-tuning on the replacements. The model pages for gpt-6-luna and the GPT-5.6 models list fine-tuning as “Not supported”. GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini supported it. Separately, OpenAI says active customers can no longer create new fine-tuning jobs from 6 January 2027. A team that relied on a fine-tuned GPT-4o has to rebuild that behavior with prompts, examples or retrieval, and test it.
• More context, newer knowledge. GPT-6 Luna has a context window of 1,050,000 tokens and a knowledge cutoff of 18 May 2026, against 128,000 tokens and October 2023 for GPT-4o. Prompts above 272,000 input tokens are billed at a higher long-context rate.

The new Model Spec: fewer style rules, more weight on intent

The Model Spec is OpenAI’s public description of how it wants its models to behave. The version dated 29 October 2026 reorganizes it around five values: faithfulness to human intent, minimizing harm, honesty and transparency, human agency, and care. It now calls the model “GPT” instead of “the assistant”. It adds a section called “Consider context carefully” on how memories and earlier conversations should shape what the model infers. And, in OpenAI’s words, it “retires much of the detailed style guidance”. The Spec is published in the public domain under CC0. Unite.AI reported the same points.

I compared the section headings of the new version with the previous one from 18 August 2026. Several style sections no longer appear, among them “Use Markdown with LaTeX extensions”, “Be concise and conversational”, “Be warm”, “Be suitably professional” and “Be thorough but efficient, while respecting length limits”. Some others were renamed or merged, so a heading count says little. The Spec’s overview now says it “intentionally leaves many details of style, personality, and interaction unspecified”, and that user preferences fill those gaps.

What stays matters as much. The rule for machine-read output is still there: “In programmatic settings, GPT should output exactly what the preceding message has asked for, in the exact format specified.” A JSON schema or an instruction like “just the code” signals that setting. The order of authority is still OpenAI’s root rules, then developer instructions, then user instructions; text inside tool output or pasted documents carries no authority by default. And the default to approach questions objectively stays.

The new part is how instructions get read. Under “Respect the letter and spirit of instructions”, the Spec asks the model to use its best understanding of intent, to notice when a source “may also have overlooked an important consequence”, and, when intent is unclear, to give an answer that works across plausible readings or to state its assumptions. For a chat assistant that is often what people want. For a pipeline that parses every character, a model that second-guesses a prompt is a bug.

Two cautions. OpenAI writes that “Our production models do not yet fully reflect the Model Spec.” The document is a training target, not a description of what your API calls do today. And publishing it did not change any model by itself. The practical consequence for builders, in my reading: do not rely on unwritten defaults. If you need Markdown, plain text, three bullets or a fixed tone, put it in the instructions, use a schema where you can, and test it on each new model. On a related note, OpenAI also updated GPT-6 in ChatGPT on 29 October for Free and Go users and logged-out visitors, to give “more concise responses and reduce overly flattering language”. Why style matters so much to readers is the subject of why AI text is so tiring to read.

How to find every model ID

Start with the code, then look where code search cannot see. A plain text search across all repositories catches most hard-coded IDs, including .env, YAML and JSON files. Here is a short, generic starting point:

grep -rnE --exclude-dir={.git,node_modules,dist,build} \
  'gpt-4o|gpt-4\.1|gpt-5-(nano|mini|pro)|"gpt-5"|\bo3(-pro)?\b|\bo1\b|o4-mini' .

It works with the grep on macOS and Linux. Read every match before you change anything: short names such as o1 and o3 also match things that are not models, and the plain gpt-5 is only caught in quotes, so that the newer gpt-5.6 names do not flood the results. Then check the places a search misses:

• IDs built at runtime, for example a version number glued onto “gpt-“. Search for the fragment too.
• Databases, feature flags and admin settings that store a model name per customer or per workflow.
• Libraries and tools with their own defaults, such as an SDK wrapper, an agent framework or a plugin that picks a model for you.
• Everything outside the main repo: infrastructure code, notebooks, spreadsheets with scripts, and no-code automations.
• The API itself. OpenAI’s list-models endpoint (GET /v1/models) returns an optional shutdown_date for each model. A small scheduled check that compares your configured models with that field turns a surprise into a ticket.

A migration plan you can run before the deadline

1. Fix what is already broken. Anything that still calls a model from the 23 October list is failing now.
2. Move the model name into one place. One setting per use case, not forty copies in forty files. It makes the switch, and a rollback, a one-line change.
3. Build a small golden set from real traffic. Fifty to a few hundred representative requests per use case, with sensitive data removed, plus the answers you consider good. If your tests live in OpenAI’s Evals platform, export them before it shuts down on 30 November. OpenAI’s own cookbook describes a move to the evaluation tool Promptfoo.
4. Run old and new side by side while you still can. Until 14 January, GPT-4o still answers. Compare format compliance (does the JSON parse, are the headings where your front end expects them), tool calls, refusals, answer length, latency and cost per task, not per token.
5. Test the prompts that leaned on old style. Look for prompts that never stated a format because the old model “just did it”: Markdown answers, a friendly tone, short replies. Write those expectations into the instructions and test again.
6. Choose the reasoning effort deliberately. Try two or three levels on the golden set and pick the cheapest one that passes. Set max_output_tokens with room for reasoning.
7. Roll out in steps. Send a small share of traffic to the new model, watch errors and cost, then widen it. Keep the rollback ready.
8. Stop the old IDs from coming back. Add the search pattern to your CI pipeline so a copy-pasted example does not reintroduce a retired model.

What this means for companies

The direct cost is an engineering sprint, not a strategy change. The indirect costs hide in three places. Fine-tuned models: whoever paid to train GPT-4o for their tone or their forms loses that work in January and cannot fine-tune the listed replacement. Budgets: teams on gpt-5-mini or gpt-5-nano face higher prices per token in December unless they test cheaper options, while teams on GPT-4o may save money, as long as reasoning stays modest. And testing: the evals platform many teams used for exactly this kind of comparison closes first. The broader lesson is about dependency. A hosted model is a product with an end date. Companies that keep their model names in one place, own their test sets and watch the deprecations feed treat a retirement like a library update. Everyone else finds out from an error log.

What this means for users

If you use ChatGPT directly, this round changes little: GPT-4o and GPT-4.1 already left ChatGPT in February, and today’s models there are newer. What you may notice is in other products. Customer-service bots, writing tools and apps that ran on GPT-4o behind the scenes will switch models around the deadlines. Answers may become shorter or longer, more careful or more direct, and some features may pause while a provider migrates. If a tool you depend on suddenly behaves differently in December or January, that is a likely reason; ask the provider which model it now uses. The new Model Spec also says style should follow your preferences, so it is worth stating them plainly: “answer in three sentences”, “no bullet points”, “be blunt”. For what you get on the paid ChatGPT plans right now, see what ChatGPT Pro’s doubled limits actually give you.

Limits of this guide

• Dates and prices come from OpenAI’s pages as of 1 November 2026. OpenAI can change them, and the pricing page notes promotional terms. I did not see the emails OpenAI says it sent to affected customers.
• The trigger was a developer blog post. A post on DEV Community on 31 October flagged the retirements and suggested a search pattern. I used it as a tip and checked every date against OpenAI’s deprecations page; the search pattern above is my own, simpler version.
• I did not measure behavior. I did not run GPT-4o against GPT-6 Luna for this article. Whether the replacement is better, worse or just different for your task depends on your prompts, which is why the plan above starts with a test set.
• The Model Spec describes a goal. OpenAI says its models do not yet fully follow it. The heading comparison is my own, between the HTML versions of 18 August and 29 October 2026.
• The notice-period question is open. The page promises six months for generally available models with exceptions, and gives three months for the GPT-4o family without saying why. I asked no one at OpenAI for this article; the gap is stated, not explained.

Related reading on ThorstenMeyerAI.com

• GPT-6.1 Terra is cheap per task. In our newsroom test, its max setting was the worst
• The AI price war: OpenAI and Anthropic slash prices. Read the fine print
• OpenAI doubles ChatGPT Pro limits. What you actually get now
• Why AI text is so tiring to read: six patterns, and how to fix them

Sources

• OpenAI, Deprecations (API documentation; entries of 22 April, 11 June, 14 October 2026 and others; checked 1 November 2026)
• OpenAI, Pricing (standard, short-context prices; checked 1 November 2026)
• OpenAI, model pages for GPT-6 Luna, GPT-4o, GPT-4.1, GPT-5-nano and GPT-5.6 Luna (checked 1 November 2026)
• OpenAI, Using GPT-6: migration quickstart and Reasoning models (API guides)
• OpenAI, An updated Model Spec for more capable and beneficial AI (29 October 2026)
• OpenAI, Model Spec, version of 29 October 2026, the previous version of 18 August 2026 and the changelog
• OpenAI Help Center, Model release notes (GPT-4o retirement in ChatGPT, 29 January and 13 February 2026; GPT-6 in ChatGPT update and Model Spec update, 29 October 2026)
• Unite.AI, OpenAI reorganizes Model Spec around five core behavior values (October 2026)
• DEV Community, OpenAI is retiring GPT-4o, 4.1 and 5-nano: how to find every model ID in your repo (31 October 2026; used as a tip)

Researched and drafted with AI assistance, including Anthropic’s Claude.