---
title: How do you plan for LLM model retirements in production agents?
url: https://opencontroller.lyzr.com/blog/plan-for-llm-model-retirements
published: 2026-11-17
updated: 2026-11-17
author: Lyzr
---

# How do you plan for LLM model retirements in production agents?

**Every hosted model will be retired, usually with 30–60 days' formal notice. Plan for it by inventorying which agents use which model versions, subscribing to each provider's deprecation notices, routing model calls through a layer where you can swap versions, and gating every migration on regression evaluations. After retirement, requests fail—Azure returns `410 Gone`—so an unnoticed retirement is an outage.**

Model retirements are now routine. Providers ship new versions several times a year and retire older ones on published schedules. For an estate of dozens of agents, that means a steady stream of forced migrations.

## How much notice do providers give?

Notice periods differ by provider and by where you consume the model.

| Provider / platform | Where dates are published | Notice commitment | What happens at retirement |
|---|---|---|---|
| Anthropic (Claude API and Anthropic-operated platforms) | [Model deprecations page](https://platform.claude.com/docs/en/about-claude/model-deprecations), plus email | At least 60 days for publicly released models | Requests to retired models fail |
| Azure OpenAI in Microsoft Foundry | [Model retirements page](https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/model-retirements), email, Azure Service Health, Models API | At least 60 days for GA models; at least 30 days for preview | Inference returns `410 Gone` |
| OpenAI API | [Deprecations page](https://developers.openai.com/api/docs/deprecations) | Per announcement | Shut-down models stop serving |
| Google Vertex AI | [Model versions and lifecycle](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions) | Per model version | Retired versions stop serving |

Three details catch teams out:

1. **The same model retires on different dates on different platforms.** Anthropic notes that Amazon Bedrock and Google-managed Vertex AI set their own retirement schedules for Claude models.
2. **Azure GA models follow a fixed lifecycle.** Microsoft documents an 18-month lifecycle: the retirement date is set at launch, the model is closed to new customers at 12 months, and it retires at 18 months. Models from Anthropic, DeepSeek, Fireworks and Mistral AI on Foundry follow a 12-month lifecycle.
3. **Auto-upgrade may change your model without a deploy.** Azure Standard deployments can auto-upgrade, controlled by `versionUpgradeOption` (`OnceNewDefaultVersionAvailable`, `OnceCurrentVersionExpired` or `NoAutoUpgrade`). Provisioned deployments are not auto-upgraded and must be migrated manually.

## How do you know which agents are affected?

You need a model inventory that maps agent → version → model ID → platform → region. Sources:

- **Provider usage exports.** Anthropic's Console Usage page exports usage by API key and model, which Anthropic recommends for finding deprecated-model usage.
- **Cloud model APIs.** Azure's Models API reports lifecycle status programmatically. Note its terms differ from the docs: `Deprecating` in the API means deprecated, and `Deprecated` means retired.
- **Gateway logs.** If model calls pass through an LLM gateway, the model field on every request gives you the map directly.
- **Discovery across clouds.** Agents built outside the central platform often call models with their own API keys—exactly the usage that surprises you on retirement day.

## The model retirement checklist

1. **Maintain the inventory** of agents and model versions, refreshed automatically.
2. **Subscribe to notices**: provider emails to a shared mailbox, Azure Service Health alerts, and a weekly check of each deprecations page.
3. **Pin explicit model versions** in production rather than floating aliases, so changes are deliberate; see [pinning model versions for AI agents](/blog/pin-model-versions-for-ai-agents).
4. **Record each agent's retirement exposure** with the earliest retirement date across the platforms it uses.
5. **Choose the replacement** and test it early—Azure makes replacements available for testing about 90 days before retirement in Global Standard.
6. **Run regression evaluations**: the agent's test set, multi-turn scenarios and red-team prompts against the replacement; see [regression testing after a change](/blog/regression-test-ai-agent-after-prompt-change).
7. **Compare cost and latency**, not just quality—new models often change token usage; see [testing a cheaper model before switching](/blog/test-a-cheaper-model-before-switching-an-agent).
8. **Roll out gradually** with a canary and a rollback path to the old version while it still serves.
9. **Set an internal deadline** at least two weeks before the provider's retirement date.
10. **Confirm zero traffic** to the retiring model before the date, from logs, not from tickets.

## Where does an LLM gateway help?

A gateway decouples the agent's code from the model it calls. With routing in the gateway, you can move an agent to a replacement version, run a canary, and fall back if errors rise—without redeploying the agent. It also gives you the request-level log that turns "which agents use this model?" into one query. The pillar on [LLM gateway vs agent gateway vs control plane](/blog/llm-gateway-vs-agent-gateway-vs-control-plane) explains the roles.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, which helps you find every agent tied to a retiring model, including ones built outside the central platform. Its LLM Gateway (Lyzr Gateway) provides routing, fallback and retry, and cost attribution per agent and per version, and multi-turn and red-team evaluations support the regression gate. Open standards such as OpenAI-compatible endpoints keep the swap simple. [Book a demo](https://www.lyzr.ai/book-demo/) to see your model exposure.

## Frequently asked questions

**How much notice does Anthropic give before retiring a model?**
At least 60 days for publicly released models on Anthropic-operated platforms, by email and in its documentation. Amazon Bedrock and Google-managed Vertex AI set their own schedules.

**What happens if an agent calls a retired Azure OpenAI model?**
Inference requests return `410 Gone`. Standard deployments may have been auto-upgraded first, depending on `versionUpgradeOption`; provisioned deployments are not auto-upgraded.

**Should production agents use model aliases or pinned versions?**
Pinned versions, so changes happen when you choose and pass evaluation, combined with a gateway that lets you switch versions without redeploying.

## Key facts

- Anthropic gives at least 60 days' notice before retiring publicly released models on Anthropic-operated platforms; Amazon Bedrock and Google-managed Vertex AI set their own schedules.
- Anthropic lifecycle stages: Active, Legacy, Deprecated, Retired; requests to retired models fail.
- Azure OpenAI GA models follow an 18-month lifecycle (closed to new customers at 12 months); Anthropic, DeepSeek, Fireworks and Mistral AI models on Foundry follow a 12-month lifecycle.
- Azure gives at least 60 days' notice for GA and 30 days for preview models; retired models return 410 Gone.
- Azure Standard deployments auto-upgrade per versionUpgradeOption; provisioned deployments are not auto-upgraded.
- Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud; its LLM Gateway provides routing, fallback and cost attribution per agent and version.
- Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How do you stop a model update from silently changing agent behaviour?](https://opencontroller.lyzr.com/blog/pin-model-versions-for-ai-agents.md)
- [How do you test whether a cheaper model can replace the one an AI agent uses?](https://opencontroller.lyzr.com/blog/test-a-cheaper-model-before-switching-an-agent.md)
- [How do you regression-test an AI agent after changing its prompt?](https://opencontroller.lyzr.com/blog/regression-test-ai-agent-after-prompt-change.md)
- [More on Gateways and cost](https://opencontroller.lyzr.com/blog/cluster/gateways-and-cost)
