AI Model Shutdowns Are Announced Months Ahead. Who in Your Company Reads Them?
AI Tools & Software

AI Model Shutdowns Are Announced Months Ahead. Who in Your Company Reads Them?

7 min read
Share:

On Thursday, 19 November, Microsoft retired seven Azure OpenAI models: o1, o1-pro, o3, o3-deep-research, o3-pro, o3-mini and o4-mini. Under Microsoft's lifecycle policy, a retired model is removed from service and every request to it returns 410 Gone. Some deployments were switched to a newer model automatically. Others simply stopped answering.

None of this was a surprise to anyone who reads vendor documentation. The dates had been published for months, and more are coming: 30 November for Claude Sonnet 4.5 and three OpenAI developer tools, 11 December for GPT-5 and o3 on OpenAI's own API, and 11 January for a specific EU data residency option. The question for a company with 100 or 1,000 people is simpler and less technical: does anyone inside the business know which model their AI integrations call, and when that model goes away?

The dates between now and mid-January

The table below covers the items most likely to sit inside a mid-size company's integrations. It is not a complete list; every vendor publishes its own.

DatePlatformWhat stopsVendor's suggested replacement
19 Nov 2026 (passed)Azure OpenAI in Microsoft Foundryo1, o1-pro, o3, o3-deep-research, o3-progpt-5.6-sol
19 Nov 2026 (passed)Azure OpenAI in Microsoft Foundryo3-mini, o4-minigpt-5.6-terra
30 Nov 2026Claude API, Microsoft Foundry and other Anthropic-operated platformsClaude Sonnet 4.5claude-sonnet-5-5
30 Nov 2026OpenAI APIAgent Builder, the Evals dashboard and API, reusable prompt objects (/v1/prompts)Agents SDK or ChatGPT Workspace Agents; Promptfoo; prompts kept in your own code
11 Dec 2026OpenAI APIgpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3, o3-progpt-6.1-sol, gpt-6.1-terra, gpt-6-luna
11 Jan 2027OpenAI APIGPT-5.2, GPT-5.4 and GPT-5.4 mini with EU data residency on the fast or priority tierSame models on standard processing, or the GPT-6.2 family

Sources: Microsoft's model retirement schedule, OpenAI's deprecations page and Anthropic's model deprecations page. Anthropic's dates apply to the platforms it operates; Amazon Bedrock and the Google-managed listing on Google Cloud set their own schedules.

The loud failure and the quiet one

A retirement reaches a business in one of two ways, and they need different kinds of attention.

The loud way is an error. If a deployment on Azure is set to NoAutoUpgrade, Microsoft's deployment documentation says it stops working on the retirement date and the code has to be pointed at a different deployment. The same happens with any call to a shut-down model on OpenAI's or Anthropic's API. An invoice-reading step, a ticket classifier or a quote assistant just fails. Someone notices within hours because a process stops, and the fix is usually a day of developer time, assuming a developer is available and the original builder left notes.

The quiet way is the automatic upgrade. On Azure, deployments set to upgrade once the current version expires are moved to the vendor's default replacement when the old model retires. Nothing breaks, so nobody looks. But the model answering your requests is now a different one. It may format output differently, follow instructions more or less literally, write longer answers and bill at a different rate. Check what the replacement costs under your own agreement before the switch, not after the next invoice.

Anthropic's documentation shows a third, smaller trap: parameters can disappear too. From API version 2026-10-21, its API rejects temperature, top_p and top_k with a 400 error, so a migration can fail on a setting that was never a problem before.

My read is that the quiet failure is the more expensive one for a mid-size company. A broken integration gets fixed because it hurts. A slightly worse or slightly more expensive one can run for months, and when somebody finally asks why the AI assistant "got worse in November", the link to a vendor's retirement table is hard to make.

The EU detail that is easy to miss

Many European companies chose EU data residency for their OpenAI usage for GDPR reasons. On 18 November OpenAI gave notice that GPT-5.2, GPT-5.4 and GPT-5.4 mini, when used with EU data residency on the fast or priority tier, will be deprecated on 11 January 2027. The suggested path is the same models on standard processing or the newer GPT-6.2 family, according to the deprecations page.

This is narrow, and most companies will not be affected. But it shows something general: the configuration you picked for compliance reasons can have a shorter life than the default one. Whoever owns the integration should know which tier and which region it runs on, and not just which model.

The test tool is on the list as well

One line in OpenAI's table deserves a second look. On 30 November the Evals dashboard, API and graders shut down, and OpenAI points users to Promptfoo, an open-source testing tool. Evals is the feature many teams used to check whether a new model still gives acceptable answers. If your developer or vendor relied on it to approve model changes, the safety net disappears on the same day as several migrations become due.

The lesson is not specific to OpenAI. Tests that live inside one vendor's console belong to that vendor's roadmap. A set of real examples with the expected answers, kept in your own repository or even a shared spreadsheet, survives any vendor change.

A hypothetical example

Imagine a 180-person logistics company in Ruse. In spring 2025 an external developer built a small service on Azure that reads incoming customer emails, extracts the shipment number and sorts each message into one of six queues. It ran on o4-mini, with the deployment left on automatic upgrades. The developer moved on to other projects; the operations manager knows the tool as "the email sorter".

On 19 November the deployment moved to gpt-5.6-terra. Nothing failed. Over the following weeks the team notices that more emails land in "urgent", and finance sees a higher Azure line on the December invoice. Both have explanations, but nobody connects either one to a model change, because nobody in the company knew there had been one.

Nothing in this scenario requires a mistake by anyone. It only requires that the model name lived in the developer's head and nowhere else.

One question to send this week

Vendors give long notice. OpenAI promises at least six months for generally available models, though preview models can go with as little as two weeks. Anthropic gives at least 60 days. Microsoft states that retirement dates are not extendable. The information exists; the gap is inside the company.

So the practical step is a single message to whoever built or runs each AI integration, whether that is an internal developer, an agency or a SaaS vendor: which model and version does it call, on which platform and region, and what is the shutdown date? Ask them to add what happens on that date: automatic switch, error, or manual migration. Anthropic customers can answer part of this from the Console usage export, which breaks usage down by API key and model; OpenAI's model list exposes a shutdown_date field for the same purpose.

Then put each date in the calendar of a named person, 30 days early, together with the twenty or so real examples that should be re-run before the switch. That is the whole discipline of running AI after go-live, and it is cheaper than one week of a quietly misrouted inbox. If no one in the company is in a position to own this, it is the kind of work our AI Ops service covers, but the question above is worth asking either way.

Frequently asked questions

Which AI models were retired on 19 November 2026?

Microsoft retired seven Azure OpenAI models in Microsoft Foundry: o1, o1-pro, o3, o3-deep-research and o3-pro, with gpt-5.6-sol as the suggested replacement, and o3-mini and o4-mini, with gpt-5.6-terra as the replacement. Requests to a retired model return a 410 Gone error.

What happens to an Azure OpenAI deployment when its model retires?

It depends on the deployment's upgrade setting. Deployments set to upgrade when the current version expires are moved automatically to the default replacement. Deployments set to NoAutoUpgrade stop working on the retirement date, and the code has to be pointed at another deployment. Microsoft says retirement dates cannot be extended.

What other AI model shutdowns are coming before January 2027?

On 30 November 2026 Anthropic retires Claude Sonnet 4.5 on the platforms it operates, and OpenAI shuts down Agent Builder, the Evals dashboard and API, and reusable prompt objects. On 11 December OpenAI shuts down gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3 and o3-pro on its API. On 11 January 2027 GPT-5.2, GPT-5.4 and GPT-5.4 mini with EU data residency on the fast or priority tier are deprecated.

Why can an automatic model upgrade be a problem?

Nothing breaks, so nobody checks. The new model can format answers differently, follow instructions differently, produce longer output and bill at a different rate, which can quietly change both the quality and the cost of an integration.

How can a company find out which AI models its integrations use?

Ask whoever built or runs each integration which model and version it calls, on which platform and region, what the shutdown date is and what happens on that date. Anthropic's Console usage export shows usage by API key and model, and OpenAI's model list includes a shutdown_date field.

Related Articles

ChatGPT for Sales Can Update Your CRM. Set the Write Rules First

OpenAI's ChatGPT for Sales lets sellers prepare calls and update CRM records from a chat. HubSpot's own documentation says its custom validation rules don't apply to those updates, so decide what ChatGPT may write before anyone installs the plugin.

Nov 19, 2026

Microsoft's Autopilot Agent Decides for Itself When to Ask You

Microsoft's new Autopilot agent in Copilot works without prompts and asks for approval only when its own check finds an action outside your request. It is US-only for now, but its rules show what companies in Bulgaria should prepare: accountable owners, tighter briefs, spending limits and a GDPR review.

Nov 18, 2026

Copilot, ChatGPT or Claude: Microsoft's Cost Study Is Really About the Meter

A Microsoft-funded study puts Copilot at $30 a month for everyday users against $81 for ChatGPT Enterprise and $205 for Claude Enterprise. The bigger lesson is where each vendor's usage meter sits, why the gap narrows for agentic work, and how to run the same test on your own tasks.

Nov 17, 2026