On Thursday, October 23, OpenAI switches off gpt-3.5-turbo, gpt-4, gpt-4-turbo, o1, o3-mini, o4-mini, and a handful of others. Any automation that still calls those names stops getting answers. It’s the biggest of ten dated shutdowns and price changes between now and January 20, and most of them hit model names that someone wired into a Zapier step or a script two years ago and hasn’t looked at since.
Here’s every date we could confirm on the vendors’ own pages, what it breaks, and what to do about it.
The calendar
| Date | What changes | What to do |
|---|---|---|
| Oct 23 | OpenAI shuts down gpt-3.5-turbo, gpt-4, gpt-4-turbo, gpt-4o-2024-05-13, gpt-4.1-nano, o1, o1-pro, o3-mini, o4-mini, and gpt-image-1, plus fine-tunes built on several of them. |
Search for these names now. Two days isn’t much time. |
| Oct 31 | DeepSeek starts charging $0.75 per 1,000 calls for its Search and Fetch tools. | Check whether an agent of yours calls them. |
| Nov 19 | Azure OpenAI retires o1, o1-pro, o3, o3-pro, o3-mini, and o4-mini. Standard deployments upgrade automatically. Provisioned (PTU) deployments just stop. |
If you pay for provisioned throughput on Azure, plan the move now. |
| Nov 30 | Anthropic retires Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) on its API. OpenAI shuts down Evals, Agent Builder, and the reusable Prompts API. |
Test on Claude Sonnet 5.5 first: it rejects two settings older code often sends. Move any stored OpenAI prompts into your own code. |
| Dec 1 | Gemini 4 Argon’s launch price ends. It goes from $2 to $4 per million input tokens and from $10 to $20 per million output. OpenAI also retires three older image models. | Budget for double the cost if you built on Argon this fall. |
| Dec 11 | OpenAI shuts down the original GPT-5 snapshots (gpt-5-2025-08-07 and its mini, nano, and pro versions) and the dated o3 and o3-pro snapshots. |
Code that pinned a dated GPT-5 version needs a new pin. |
| Jan 1 | Gemini 3.6 Flash and 3.8 Flash double to $1.50 and $7.50 per million tokens. | Re-run your cost math for 2027. |
| Jan 6 | OpenAI retires its classic text-to-speech models (tts-1, tts-1-hd) and stops accepting new fine-tuning jobs from existing customers. |
Move voice features to the newer realtime models. |
| Jan 14 | OpenAI shuts down the GPT-4o and GPT-4.1 snapshots: 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. |
This is the one most small businesses will feel. See below. |
| Jan 20 | OpenAI retires its first-generation realtime and audio models, including gpt-realtime, gpt-4o-realtime, and gpt-audio. |
Phone bots and voice agents built in 2025 need an upgrade. |
The twenty-minute check: search your code, environment variables, and every Zapier, Make, or n8n step for gpt-3.5, gpt-4, o1, o3, o4-mini, and claude-sonnet-4-5. Anything you find is on the calendar above. If you call the plain names gpt-4o or gpt-4o-mini, OpenAI hasn’t said whether those aliases follow their snapshots in January. Its models endpoint reports a shutdown_date for each name, so ask it rather than guess.
Pick the replacement on price, not on the default
OpenAI’s own shutdown table points the October 23 models at its GPT-5.6 family. But on October 7 OpenAI started recommending the newer GPT-6 family for production, and the cheapest member, GPT-6 Luna, costs $0.10 per million input tokens and $0.50 per million output. Moving a gpt-4o step to Luna cuts its token bill by more than 90 percent. Claude Haiku 5.5, released October 7, sells at the same price.
Cheap isn’t automatically right. Run twenty real examples through the replacement and compare the answers before you switch. And read the fine print, because long prompts now cost extra almost everywhere: Haiku 5.5 bills five times its base rate once a prompt passes 100,000 tokens, GPT-6 models cost more past 272,000, and Grok doubles at 200,000. Our LLM Ranking Tool has current prices for 28 models if you want to compare before you test.
Four other changes worth an hour this month
Flat seats are turning into seats plus credits. Microsoft is moving advanced Copilot features to usage-based Copilot Credits starting this quarter. Nothing bills until an admin creates a spending policy, so create a small one on purpose rather than a large one by accident. GitHub Copilot overage now runs at a penny per credit and is on unless an admin turns it off. ChatGPT Business features like its Word and PowerPoint add-ins draw from a metered credit pool.
Illinois employers owe a notice. HB 3773 has been in force since January 1. If AI plays a part in hiring, promotion, discipline, or firing, you have to tell applicants and employees, and you can’t let the tool discriminate even by accident (zip code as a proxy counts). The state withdrew its draft notice rules in June, so there’s no official template. The duty still applies. Write a plain notice now and ask your HR software vendor for its bias-testing records. Colorado’s version starts January 1, 2027, and California’s SB 947 follows that July.
Your AI assistant is now a phishing target. Barracuda reported fake invoices this month with hidden instructions aimed at AI inbox assistants, telling them to suggest a change to vendor payment details. Anthropic’s September threat report says stolen AI keys are now prized loot. No AI summary or agent should approve a payment or a bank-detail change, and every API key you own should have a spending cap.
Running automations got much cheaper. Between the new budget models and the API credits now included with Claude Max and Team plans, a workflow that didn’t pencil out in the spring might now. If one got shelved over cost, price it again.
If you aren’t sure what your automations call, that’s a one-day audit for us, and the October 23 deadline is two days out. Tell us what’s running or call (630) 548-5614.