Operational AI Governance

Your Model Inventory Doesn't Know When Your Model Retires

AI providers publish retirement dates for their models. Your inventory should hold them, along with the channel you buy through.

Christopher R. McCraleyThe Decision Terrain2 min read

January 14, 2027. That's when OpenAI says several GPT-4o and GPT-4.1 model snapshots will shut down on its API, according to a notice posted on October 14.

If one of those models sits behind a customer-facing workflow at your firm, the question isn't whether the provider gave notice. It did. The question is whether anyone at your firm will see it in time to act.

What the providers publish

OpenAI's deprecations page lists several retirements announced between July and October 2026, covering audio, transcription and text models. It states minimum notice periods: at least six months for generally available models, at least three months for specialized variants, and much shorter notice for preview models.1

Anthropic's policy is at least 60 days' notice for publicly released models. On September 30, 2026, it announced that Claude Sonnet 4.5 will retire on November 30, 2026, on its own platforms. Its page adds that partner-operated platforms, Amazon Bedrock and Google Cloud, set their own retirement schedules, so dates can differ.2

The same model can retire on different dates depending on where you buy it. An inventory that records only the model's name can't tell you which date applies. If you reach a model through a cloud marketplace, that platform's table, not the provider's, holds your date.

The fields that are missing

Model: [name, version]
Bought through: [provider API, cloud platform, embedded in a vendor's product]
Retirement date on that channel: [date, with the source link]
Minimum notice on that channel: [from the provider's policy]
Our time to revalidate a replacement: [weeks to test and approve]
Alert set for: [retirement date minus revalidation time]

Compare the fifth line with the fourth. If it takes your firm twelve weeks to test and approve a replacement and the channel promises sixty days' notice, you will always start late. Either shorten the revalidation or start it before the notice arrives. And record where the replacement will come from before you need it.

Models embedded in a vendor's product are the hardest case. The vendor gets the provider's notice. Do you get the vendor's? Put that in the contract: notice of the underlying model's retirement, and of the replacement, before it happens. Then treat the date like a contract renewal: owned, on a calendar and reviewed in advance.

This is different from a provider silently updating a model you approved. A retirement is announced and dated. The risk isn't surprise. It's that the date reaches an inbox nobody connects to the workflow it will break.

For a startup whose product rests on one model snapshot, this is a roadmap item, not an IT task. Put the date on the product calendar.

The provider published the date. The rest of the work is getting it to the person who has to act.

Read Five Substitutes for a Plan for what organizations actually rely on when a critical process stops.

Sources and Notes
  • 1. OpenAI, "Deprecations," API documentation, checked October 18, 2026. Supports the October 14, 2026 notice of GPT-4o and GPT-4.1 snapshot shutdowns on January 14, 2027, other retirements announced July 20 to October 1, 2026, and the stated minimum notice periods. It is the provider's own documentation of its policy.
  • 2. Anthropic, "Model deprecations," Claude Platform documentation, checked October 18, 2026. Supports the minimum 60 days' notice for publicly released models, the September 30, 2026 announcement of Claude Sonnet 4.5's retirement on November 30, 2026, and the statement that partner-operated platforms set their own schedules. It is the provider's own documentation. Neither source compares providers or addresses customers' revalidation times. The inventory fields, the notice-versus-revalidation comparison and the embedded-vendor clause are the author's analysis; naming these providers is not a comparison or endorsement.
About The Decision Terrain

The Decision Terrain — Driving Integrity in the Era of AI

The Decision Terrain examines how organizations preserve judgment, evidence, accountability, and the capacity to act when they must make critical decisions under stress.

All articles in The Decision Terrain are authored by Christopher McCraley.

About the author
Christopher R. McCraley, Founder and Principal Advisor, SRS Platinum.
Christopher R. McCraley
Founder & Principal Advisor

Founder and Principal Advisor of SRS Platinum, Inc., an institutional advisory practice focused on operational resilience, AI governance, infrastructure modernization, and operational decision integrity. Four decades across financial services, insurance, infrastructure, consulting, and enterprise technology environments.

How to cite this article

McCraley, C. R. (2026). Your Model Inventory Doesn't Know When Your Model Retires. The Decision Terrain, SRS Platinum, Inc. https://srsplatinuminc.com/ai-model-retirement-inventory/

The views expressed in this article are solely my own and do not represent the views, positions, or policies of any current or former employer, client, or affiliated organization. Any professional examples are anonymized and presented solely to illustrate broader principles.

Scroll to Top