Every AI model your business runs on has a shutdown date. Several land between now and February. Here is the calendar, how each vendor handles notice, and how to switch models without breaking the work.
November 30, 2026.
That is the date Anthropic has scheduled to retire Claude Sonnet 4.5 from its API. After that, any application still calling it gets an error instead of an answer. Anthropic marked the model deprecated on September 30 and named Claude Sonnet 5.5 as the replacement.
It’s not alone. Between now and the end of February, OpenAI, Anthropic, Microsoft, and Google are retiring a long list of models that real businesses built real workflows on. The quoting assistant in your sales team’s browser, the document summarizer your operations manager set up last spring, the chatbot on your website: all of them run on a specific model, and all of those models have an expiration date.
Most companies have no idea which dates apply to them.
The Shutdown Calendar, November Through February
Here’s what the vendors’ own published schedules show as of mid-November.
Anthropic. Claude Sonnet 4.5 retires November 30, with Claude Sonnet 5.5 as the recommended replacement. Anthropic also removed the temperature, top_p, and top_k settings in its October 2026 API version. Requests on that version that still include them return an error. So it’s not just models that retire. Settings do too.
OpenAI. GPT-5, GPT-5 mini, GPT-5 nano, and o3 shut down in the API on December 11. OpenAI posted that notice back on June 11 and points customers to GPT-6.1 Sol, GPT-6.1 Terra, and GPT-6 Luna. Then on January 14, 2027, the API versions of GPT-4o, GPT-4o mini, GPT-4.1, and GPT-4.1 mini go away, with GPT-6 Luna as the replacement. That notice went out October 14, about three months ahead.
Microsoft. Azure customers run on a separate calendar. Microsoft’s schedule retires o3 and o4-mini, along with several other o-series models, on November 19, and lists the August 2025 GPT-5 models for February 9, 2027.
Google. Google lists retirement dates for several Gemini API preview models as the “earliest possible” dates, with exact dates to be announced. It is also limiting new access to Gemini 2.5 Pro and Gemini 2.5 Flash.
That’s four vendors and dozens of models in roughly 90 days.
Same Model, Different Dates
Here’s the detail that catches people. The same model can retire on different dates depending on where you buy it.
Claude Sonnet 4.5 leaves Anthropic’s API on November 30, but trackers show it running on Amazon Bedrock until April 2027. OpenAI’s GPT-5 leaves OpenAI’s API on December 11, while Microsoft lists it on Azure until February 9. Microsoft’s own schedule shows some Claude models leaving its Foundry platform earlier than Anthropic’s dates.
So “we use GPT-5” isn’t enough information. You need to know which GPT-5, through which platform, on which version.
How Each Vendor Handles Notice
The notice policies differ more than most buyers realize.
OpenAI promises at least six months for generally available models and at least three months for specialized variants. Preview models can go with much shorter notice, which OpenAI’s page describes as potentially two weeks. Fine-tuned models follow their base model’s date.
Anthropic commits to at least 60 days’ notice for publicly released models and emails customers with active deployments.
Google’s Gemini API publishes earliest-possible dates and fills in exact dates later. Google Cloud states longer commitments for stable models on its Vertex AI platform.
Microsoft announces replacements ahead of retirement on Azure, and its documentation is blunt: retirement dates aren’t extendable.
None of these policies is wrong. They’re different. Which one fits you depends on how fast your team can test and ship a change, and that’s a question about your process, not the vendor.
Why a Model Swap Isn’t a Simple Upgrade
It’s tempting to think of a new model as a free upgrade. Newer, smarter, done. In practice, swapping models changes behavior in ways that matter.
A newer model may write longer answers or shorter ones. It may format output differently, which breaks the code that reads it. It may refuse requests the old model handled, or handle requests the old model refused. It may call tools and connected systems differently. It may cost more or less per request, and it may respond faster or slower.
Prompts are tuned to a specific model, often by trial and error. Move that prompt to a new model and some of that tuning stops working.
What This Looks Like Inside a Real Business
Walk through a common setup. A Fox Valley manufacturer built a quoting assistant on GPT-4o in 2024. It reads a customer’s emailed request, pulls part numbers, checks them against a price list, and drafts a quote for a salesperson to review.
The tool works well. The developer who built it moved on. No one on staff has looked at the code in a year.
On January 14, the API stops answering for GPT-4o. The quoting tool doesn’t fail gracefully. It just stops producing drafts, and the sales team starts building quotes by hand again during the busiest month of the year.
Or worse, someone swaps in GPT-6 Luna in a hurry. The quotes come back, but the new model formats part numbers differently, and three quotes go out with the wrong items before anyone notices.
Neither outcome is the model’s fault. Both are planning failures.
This matters in Wisconsin right now. In October, the state announced $6.6 million in WisTRAIN grants to 84 employers to train 3,823 workers in advanced manufacturing and AI. That’s a real investment in people learning AI tools. Those tools sit on models that will change underneath them, and the training only pays off when the tools keep working.
A Plan for Model Retirements
Here’s a practical approach for any business running AI in production.
Inventory every place you call a model. List every application, script, and automation, along with the exact model name, version, and platform. Include AI features inside the software you buy. Ask your vendors which models they run on and how they handle retirements.
Pin specific versions. Many vendors offer aliases that point to “the latest” model. Those aliases can change behavior without warning. Pin a dated version so changes happen when you choose.
Build a test set. Collect 50 to 100 real examples with known good answers. When a replacement model arrives, run the full set on both models and compare the results side by side.
Test early. Start testing the replacement the week a deprecation notice lands, not the week before shutdown. Check output quality, format, cost, and speed.
Make the model a setting, not a rewrite. Build your applications so switching models is a configuration change. A thin layer between your code and the vendor makes migrations faster and lets you compare vendors honestly.
Assign an owner. Someone on your team should subscribe to every vendor’s deprecation notices and keep a calendar of upcoming dates.
Budget for it. Model migrations are now a recurring cost of running AI, like software updates. Plan for at least one or two a year per application.
The Bottom Line
No AI vendor is going to stop retiring models. OpenAI, Anthropic, Microsoft, and Google are all moving fast, and older models get shut down to make room. That’s the deal when you build on a platform you don’t own.
The businesses that handle it well aren’t the ones that picked the “right” vendor. They’re the ones that know what they run, test before they switch, and treat migration as normal maintenance instead of an emergency.
Earthling Interactive helps businesses in Madison, across Wisconsin, and throughout the continental U.S. build AI systems that survive vendor changes. As an AI consulting and custom software development partner, we inventory what you’re running, build the test sets and model layers that make switching safe, and handle migrations before the deadline does it for you. Got a model on the shutdown list? Let’s talk this week.
