The Next OpenAI API Shutoff: GPT-5.1, GPT-5.3-Codex and GPT-5.4 Nano End April 1, 2027

The next OpenAI API shutoff is set for April 1, 2027, when gpt-5.1, gpt-5.3-codex and gpt-5.4-nano stop working. OpenAI added the entry to its deprecations page on October 1, 2026, and named GPT-6 Sol and GPT-6 Luna as the replacements.

This is the wave after the one most teams are busy with now. We covered the October 23 shutoff of seventeen models on Wednesday. This post covers the April OpenAI API shutoff: how to find out whether you call these models, where to move, what it costs, and a month-by-month plan. Facts are as of October 6, 2026.

The short version: six months is enough time, and it is easy to waste. GPT-5.4 nano workloads get cheaper on GPT-6 Luna. GPT-5.1 workloads pay more for input on GPT-6 Sol. The bigger cost risk is a default setting: the old models reason at none unless told otherwise, and the GPT-6 models default to medium. Set effort on purpose, or your bill moves without anyone touching the code.

What the OpenAI API shutoff covers

The OpenAI deprecations page lists three model IDs under the October 1 entry. All three share one shutdown date.

The table names bare IDs only. The model pages also mark the dated snapshots as deprecated: gpt-5.1-2025-11-13 and gpt-5.4-nano-2026-03-17. If your code pins a snapshot, assume it goes on the same day, and search for both forms.

OpenAI’s stated policy is at least six months’ notice for generally available models. October 1 to April 1 is exactly six months, so do not expect this OpenAI API shutoff to slip.

The same October 1 update also deprecated the text-to-speech models tts-1, tts-1-hd and two gpt-4o-mini-tts snapshots, with a January 6, 2027 shutdown and gpt-realtime-2.1-mini as the replacement. If a site uses OpenAI for read-aloud audio, that date comes first. Our post on transcription model retirements for captioning covers the audio side.

Where these models hide in agency work

Each of the three tends to live in a different kind of project.

gpt-5.1 is the oldest of the three. Its dated snapshot is from November 2025, so integrations built around then may still use it. Look in chatbots on member and donor sites, content assistants in WordPress, and summary or tagging jobs in Drupal.

gpt-5.4-nano is the cheap one. It tends to run high-volume work nobody watches: alt text, spam scoring on forms, search query rewriting, metadata extraction.

gpt-5.3-codex is a coding model. It is less likely inside a client site and more likely in your own tools: code review scripts, CI jobs, or internal agents your developers set up.

The places to search are the same as in October. Plugin settings in the WordPress options table, Drupal config, environment variables in hosting dashboards, no-code tools like Zapier and Make, and serverless functions. Our October shutoff post walks through each one. If a client site is Drupal and still on version 10, note that Drupal 10 support ends December 9, 2026, well before April.

How to find out which models you call

Searching code finds the IDs people typed. It misses the ones a plugin picked for them. For the OpenAI API shutoff, the more reliable check is OpenAI’s own usage data.

OpenAI’s completions usage endpoint, GET /organization/usage/completions, can group results by model. It can also group by project_id and api_key_id. It sits in the Administration section of the API reference, so expect to need an admin key.

For an agency, grouping by API key and model together is the useful view. If each client site has its own key, you get a list of which client calls which model. Any row showing gpt-5.1, gpt-5.3-codex or gpt-5.4-nano goes on the migration list.

Two limits apply. Usage data shows only keys in organizations you control. A client who owns their own OpenAI account has to run the check, or give you access. And a model called once a month will appear only if your date range covers that call, so pull at least 30 days.

Then search code and settings for these strings:

gpt-5.1
gpt-5.1-2025-11-13
gpt-5.3-codex
gpt-5.4-nano
gpt-5.4-nano-2026-03-17

Note that gpt-5.1 also matches gpt-5.1-codex and other names that start the same way. Check each hit by hand.

What to switch to, and what it costs

OpenAI’s table names two targets. A third, GPT-6.1 Sol, launched after the GPT-6 pair and is worth weighing. Prices below are per million tokens on standard processing, from the model pages and the OpenAI pricing page.

Here is what that means for each move.

gpt-5.4-nano to gpt-6-luna: cheaper. Input is half the price and output is 60% lower. This is the easy one on cost.

gpt-5.1 to gpt-6-sol: input costs 60% more, from $1.25 to $2.00. Output stays at $10.00. A chatbot that sends long prompts and gets short answers will cost more.

gpt-5.3-codex to gpt-6-sol: input rises about 14%, and output drops about 29%. Whether it comes out cheaper depends on your mix of input and output.

Batch processing is half price on the GPT-6 models, so overnight jobs keep that saving after the move. For a wider view of the GPT-6 lineup, see our guide to which OpenAI model to use.

The default that changes your bill

This is the part of the OpenAI API shutoff most likely to surprise a client.

According to OpenAI’s model pages, gpt-5.1 and gpt-5.4-nano both default to none for reasoning effort. If your code never set an effort level, those models answered without reasoning.

gpt-6-sol and gpt-6-luna default to medium. Swap only the model ID and every request starts reasoning. Reasoning tokens bill as output, the expensive side, as we explained in input vs output tokens.

That last line matters for the choice between Sol versions. A job that needs none can move to GPT-6 Sol as a straight swap. On GPT-6.1 Sol it has to run at low at minimum, which means some reasoning cost on every call.

Context also changes. The GPT-5.x models here have a 400,000-token window. The GPT-6 models have 1,050,000. The changelog gives GPT-6 Sol prices for prompts up to 272K tokens, and longer prompts cost more. Most chatbots never get near either limit.

GPT-6.1 Sol as a replacement option

GPT-6.1 Sol launched on September 29, 2026, according to OpenAI’s API changelog. It costs the same per token as GPT-6 Sol. Cached input is half the price, $0.10 against $0.20.

That cache price matters for chatbots. A support bot sends the same long system prompt and knowledge text on every turn. With caching, most of that input bills at the cached rate, so halving the cached rate cuts a real share of the bill.

On quality, OpenAI’s GPT-6.1 Sol announcement says that on DeepSWE v1.1, GPT-6.1 Sol "matches GPT‑6 Astra at roughly one-fifth of the cost." A post from OpenAI’s developer account put the score at 75.2% at high reasoning effort, against 68.8% for GPT-6 Sol at maximum effort. These are OpenAI’s own figures on a coding benchmark. They tell you little about a member services chatbot.

There are two catches. First, the deprecations page names gpt-6-sol, not gpt-6.1-sol. GPT-6 Sol has no shutdown date, but its model page now says "See GPT-6.1 Sol for the newer Sol model," and it no longer appears on the pricing page. Second, GPT-6.1 Sol has no none effort, as noted above.

Our take: for a gpt-5.3-codex workload, go to GPT-6.1 Sol. That model never supported none, so the missing setting costs you nothing. For a gpt-5.1 chatbot that ran with no reasoning, test GPT-6 Sol at none and GPT-6.1 Sol at low side by side, and compare cost and answers. We covered why there is no GPT-6.1 Astra in OpenAI pauses GPT-6.1 Astra.

An OpenAI API shutoff migration timeline

Six months sounds long. This calendar shares space with other deadlines, so here is one way to lay it out.

1. October: finish the October 23 work first. Then pull 30 to 60 days of usage grouped by model and API key, and list every client and tool on the three IDs.
2. November: set up test runs on the replacement models. Save 50 to 100 real prompts per workload, with the answers you get today. If you use Claude, Sonnet 4.5 retires November 30, so plan that work in the same month.
3. December: compare answers, cost and speed at the effort settings you chose. The GPT-5 snapshot line shuts down December 11, so some teams will already be in this code.
4. January: move low-risk jobs first, such as alt text and tagging. Note January 6 for the `tts-1` models.
5. February: move chatbots and anything client facing. Tell clients before the switch, since answers will read differently.
6. March: confirm the old IDs show zero usage. Remove them from config. Set an alert on failed requests, not on a specific error code.
7. April 1, 2027: the OpenAI API shutoff date. Nothing should still be calling the three models.

The general fix is the same one we keep repeating. Put the model ID and effort level in one config value per site, and keep a list of which sites use what. The next deprecation notice then becomes a short task.

The same pattern at Anthropic and Google

OpenAI is not the only vendor retiring models on agency sites this season.

Anthropic retires claude-sonnet-4-5-20250929 on November 30, 2026, with claude-sonnet-5-5 as the replacement, per its model deprecations page. Anthropic’s policy is at least 60 days’ notice. A tip for your deprecation-check script: since October 1, Anthropic’s Models API returns a line field, such as opus, sonnet or haiku. Group models by line instead of parsing their IDs. We covered the breaking changes in Claude Sonnet 5.5 and the Sonnet 4.5 retirement.

Google lists gemini-3.1-flash-lite for shutdown on May 7, 2027, with gemini-3.5-flash-lite as the replacement, on its Gemini API deprecations page. The Veo 3.1 preview video models end October 22, 2026. Our post on Gemini shutdown dates this fall has the rest.

If a site uses more than one vendor, one inventory covers the OpenAI API shutoff and these dates too. Group by vendor, model and client, and sort by date.

Frequently Asked Questions