Model Sunset Desk

OpenAI API shutdown · Friday 23 October 2026

On 23 October, OpenAI shuts down gpt-4, gpt-3.5-turbo, o1, o4-mini and gpt-image-1.

The same list also has gpt-4-turbo, gpt-4.1-nano, o1-pro, o3-mini, the gpt-4o-2024-05-13 snapshot and fine-tuned models built on retiring bases. Requests that name a shut-down model fail outright. Changing the name takes a minute. The real work is finding out what the replacement does differently with your prompts.

We make the swap, test it on your own inputs, and hand you a pull request with a before/after report.

Check your code free See prices

Source: OpenAI API deprecations page, entry announced 22 Apr 2026, read 9 Oct 2026. Plain gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini are not on the 23 Oct list.

Illustrative example. Line 42 breaks loudly on 23 Oct. Lines 44 and 45 can change quietly after the swap.

Free exposure check · $0

Paste code, a config file or a workflow export. See what retires, and when.

exposure checkruns in this tab · nothing is uploaded
or press Ctrl + Enter

What breaks and what silently changes

The shutdown is loud. The swap is quiet.

Breaks on 23 Oct

Any request that names one of these IDs fails. The error is obvious, and the feature stays down until someone ships a change.

Shuts down 23 Oct 2026OpenAI recommends
gpt-3.5-turbo -0125, -completionsgpt-5.6-terra
gpt-4 -0613, -completions, -0613-completionsgpt-5.6-sol
gpt-4-turbo -2024-04-09, -completionsgpt-5.6-sol
gpt-4-1106-previewgpt-5.6-sol
gpt-4o-2024-05-13 snapshot onlygpt-5.6-sol
gpt-4.1-nano -2025-04-14gpt-5.6-luna
o1 -2024-12-17gpt-5.6-sol
o1-pro -2025-03-19gpt-5.6-sol reasoning.mode: pro
o3-mini -2025-01-31gpt-5.6-sol
o4-mini -2025-04-16gpt-5.6-terra
gpt-image-1gpt-image-2.5-sunburst or gpt-image-2.5-flare
ft-gpt-3.5-turbo ft-o4-mini-2025-04-16, ft-babbage-002, ft-davinci-002gpt-5.6-terra
ft-gpt-4gpt-5.6-sol
ft-gpt-4.1-nano-2025-04-14gpt-5.6-luna

Source: OpenAI API deprecations, 2026-04-22 entry, read 9 Oct 2026.

Can change after the swap

The replacement accepts the same request and answers differently. Nothing errors, so nothing alerts you. These are the things we measure:

Output format
Markdown where your UI expects plain text, or the reverse.
Length
Longer or shorter answers move your layout and your token bill.
Tone
A support reply that sounded like your company may stop sounding like it.
Refusals
Requests the old model handled may be declined, or the other way round.
JSON shape
A missing key, a renamed field or prose around the JSON breaks the parser downstream.
Latency
Response time can move enough to trip timeouts in webhooks and workflow steps.
Cost
Per-token prices and output length together set the new monthly bill.

Fine-tuned on a retiring base?

OpenAI’s fine-tuning guide says fine-tuned models “will remain available for inference until their base models are deprecated.” The same guide lists no GPT-5 or GPT-5.6 model for fine-tuning, and says OpenAI “is winding down the fine-tuning platform.” So you may not be able to simply retrain on the replacement. The behaviour has to be rebuilt with structured prompting, few-shot examples and output schemas, then measured against your held-out data.

Source: OpenAI fine-tuning guide, read 9 Oct 2026.

Later dates worth knowing

DateWhat changesRecommended pathSource
31 Oct 2026Existing OpenAI evals become read-onlyOpenAI links a guide to moving to PromptfooOpenAI
30 Nov 2026OpenAI Evals, Agent Builder and reusable prompts (v1/prompts) shut downAgents SDK or ChatGPT Workspace Agents; prompt content moves into your codeOpenAI
30 Nov 2026AI agent retiresAI agentAnthropic
1 Dec 2026gpt-image-1-mini, gpt-image-1.5, chatgpt-image-latestgpt-image-2.5-sunburst or gpt-image-2.5-flareOpenAI
11 Dec 2026Dated snapshots gpt-5-2025-08-07, gpt-5-mini-2025-08-07, gpt-5-nano-2025-08-07, gpt-5-pro-2025-10-06, o3-2025-04-16, o3-pro-2025-06-10gpt-5.6-sol, -terra or -luna by sizeOpenAI

All read 9 Oct 2026. Undated aliases such as plain o3 and gpt-5 are not on the 11 Dec list.

Three fixed prices · USD

Pick the job. The price and the payment terms are on the page.

Every job starts with one paragraph of scope and an acceptance test, agreed by email before any work begins. We take work with a 23 Oct deadline until our calendar is full, and we will say so before you pay.

Swap & Verify

$290per repo, or per bundle of up to 3 n8n / Make / Zapier workflows

  • Retiring IDs replaced with OpenAI’s recommended model
  • Parameter and response-shape differences fixed
  • 30-case before/after regression on your own inputs, run with your staging API key
  • Pull request (or fixed workflow JSON) plus a one-page quality and cost report
Payment
Pay on delivery. The Payment Link comes with the pull request and report. If the delivery does not pass your review, you do not pay.
Delivery
Within 48 hours of receiving access and a staging key, weekends included.
Request Swap & Verify

No payment up front. The $290 Payment Link is emailed with the delivery.

Fine-tune Replacement

$2,500for one retiring fine-tuned model

  • Your fine-tune’s behaviour rebuilt on the recommended base from your training JSONL: structured prompting, few-shot examples, output schemas
  • Held-out comparison report: accuracy, format and cost
  • Integration pull request
Payment
$1,250 deposit to start, $1,250 on delivery of the comparison report.
Delivery
3 to 5 business days.
Ask about your fine-tune

The deposit link is emailed after the scope is agreed; the balance link comes with the report.

Agency Pack white-label

$990for up to 5 client integrations

  • Swap & Verify for each client integration
  • Delivered under your agency’s brand
  • NDA on request
Payment
$495 to start, $495 on delivery.
Delivery
3 business days.
Request the Agency Pack

The deposit link is emailed after the scope is agreed; the balance link comes with the delivery.

Free: the exposure checker above, $0, instant, nothing leaves your browser.

Also available (quoted per job): moving off Agent Builder or reusable prompts (v1/prompts) before 30 Nov, and migrating from AI agent before 30 Nov.

Payment is by Stripe Payment Link (card). Prices are in USD. Full terms.

Evidence, not testimonials

We scanned the WordPress plugin directory for retiring model IDs.

scan record9 Oct 2026
source
WordPress.org plugins API, searches for “openai”, “chatgpt”, “gpt”, “ai content”, “ai chatbot”
found
1,968 unique plugins
kept
the 250 with the most active installs
method
latest stable zip, text search for exact model IDs on the 23 Oct list; nothing was executed
result
38 of 250 reference at least one retiring model ID
installs
about 3,049,000 active installs across those 38

A reference can be a hard-coded default, which breaks on 23 Oct, or an option in a dropdown, which is cosmetic. We confirmed hard-coded defaults by hand in the three plugins below.

Hard-coded defaults, confirmed by hand

  1. otter-blocks (Themeisle) · 300,000+ active installs

    inc/server/class-prompt-server.php:144

    'model' => 'gpt-3.5-turbo',

    → gpt-5.6-terra

  2. essential-blocks (WPDeveloper) · 200,000+ active installs

    includes/Integrations/AI/OpenAI.php:46

    private $model = 'gpt-4.1-nano'

    → gpt-5.6-luna · image default gpt-image-1 too (AI.php:107, OpenAI.php:353)

  3. the-post-grid (RadiusTheme) · 100,000+ active installs

    app/Controllers/Api/AIIntegration.php:61

    Default model is gpt-3.5-turbo

    → gpt-5.6-terra · its allowed-model list still includes the retired text-davinci-002/003

Source: our own scan, 9 Oct 2026. Install counts are WordPress.org’s published figures. These are open-source plugins whose vendors can ship their own fix, and may already have by the time you read this. We cite them to show how common hard-coded model IDs are. They are not our clients.

Sample deliverables you can read now

  • Sample Swap & Verify report: otter-blocks

    The patch for line 144, the regression harness, the test cases and the parameter-diff analysis. Live before/after numbers need an API key. We did not make paid API calls for this sample (budget $0), and the report says so.

  • Sample n8n workflow fix

    A workflow pinned to gpt-3.5-turbo, the fixed export, and notes on what changed and why.

  • Fine-tune replacement method

    How we rebuild a fine-tune’s behaviour on the replacement model, and the eval harness design that measures it.

How it works

Four steps. You never give us production access.

  1. Step 1

    Scope, in writing

    Tell us what calls the retiring model. We reply with one paragraph of scope and an acceptance test before any work starts. If our calendar is full, we say so before you pay.

  2. Step 2

    Access without production

    Give us access to open a pull request, or send the workflow export. Create a staging API key for this job and revoke it when we finish. We never ask for passwords.

  3. Step 3

    Swap and measure

    We switch to OpenAI’s recommended replacement, fix parameter and response-shape differences, then run your own inputs through the old and new models side by side.

  4. Step 4

    Pull request and report

    You get the pull request (or fixed workflow JSON) and a one-page report on quality, format and cost. You review it and merge it yourself.

Who we are

One operator and a set of AI coding agents.

Model Sunset Desk is one person, Orkhan Karimov, working with AI coding agents. The agents search the code, draft the patch and run the comparisons. Orkhan Karimov reviews every change before it reaches you. We call that AI-assisted, human-reviewed. It is not a team and not an agency.

This is a new business. There are no client logos or testimonials on this page because there are no past clients yet. What you can judge is the work shown here: the checker, the scan and the samples.

How every delivery is made
DrawnAI coding agents
CheckedOrkhan Karimov, every change, before delivery
Delivered asPull request or fixed workflow export
AccessStaging API key you revoke afterwards. No production access, no passwords.
Your dataUsed only for the job. Deleted within 7 days of delivery.
Rev.1 · 9 Oct 2026

Questions

Frequently asked

Is plain gpt-4o affected on 23 Oct?

No. Only the dated snapshot gpt-4o-2024-05-13 is on the 23 Oct list. Plain gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini are not on it. gpt-4.1-nano is. Source: OpenAI deprecations page, read 9 Oct 2026.

Does the checker send my code anywhere?

No. It is plain JavaScript in checker.js, running in your tab. It contains no network code, and the page’s Content-Security-Policy (connect-src 'none') blocks scripts from making connections. You can disconnect from the internet after the page loads and it still works.

Why not just change the model name myself?

You can, and for some projects that is enough. The list and the checker on this page are free. What a rename does not tell you is whether the new model still returns the format, length, tone and JSON your code expects. If you have time to run your own before/after comparison, do. If you don’t, that comparison is what we sell.

Do you guarantee the new model behaves the same?

No. Nobody can promise identical output from a different model. We promise a measured comparison on your own inputs and a pull request. We do not guarantee output parity, quality, uptime, revenue or compliance.

What access do you need?

Enough to open a pull request in your repository (or a workflow export you import yourself), plus a staging API key that you create for the job and revoke afterwards. We do not take production access or passwords.

We fine-tuned a model on a retiring base. Can’t we just retrain on the replacement?

Maybe not. OpenAI’s fine-tuning guide lists no GPT-5 or GPT-5.6 model for fine-tuning and says the fine-tuning platform is being wound down (read 9 Oct 2026). The Fine-tune Replacement offer rebuilds the behaviour with structured prompting, few-shot examples and output schemas, then compares it on held-out data from your own training set.

Can you finish before 23 Oct?

Swap & Verify is delivered within 48 hours of receiving access and a staging key, weekends included. The Agency Pack takes 3 business days and Fine-tune Replacement 3 to 5. We take work with a 23 Oct deadline until our calendar is full, and we will say so before you pay.

What about Agent Builder, reusable prompts or Claude Sonnet 4.5?

Those have later dates (30 Nov 2026). We can quote a move off Agent Builder or v1/prompts, or a migration from AI agent, separately. They are not part of the fixed prices above.

Do you have clients or reviews?

Not yet. This is a new business, so there are no logos or testimonials to show. The scan, the samples and the checker are there so you can judge the work directly.

How does payment work?

By Stripe Payment Link, by card, in USD. Swap & Verify is paid only after you have reviewed the delivery. Fine-tune Replacement and the Agency Pack are half to start and half on delivery. Services to non-residents of Canada are generally zero-rated for GST/HST. See the terms for refunds.

Contact

Write to hello@modelsunsetdesk.com

Include what calls the model (repository, workflow or plugin), the IDs the checker found, and your deadline. We reply with a scope paragraph and an acceptance test, or tell you plainly if it is not a fit or our calendar is full.