How to know when an AI model you use is retiring

Where each provider publishes retirement dates

Start with the official page for each place you call a model. Each one works a little differently.

Google Cloud's Agent Platform is what Google calls the evolution of Vertex AI (Google Cloud blog), so that's where Vertex model dates now live.

How much notice you get

• OpenAI: at least 6 months for generally available models, at least 3 months for specialized variants, and as little as 2 weeks for preview models (OpenAI).
• Claude API: at least 60 days before retirement for publicly released models (Claude deprecations page).
• Gemini API: the listed dates are the earliest a model might retire, and Google says it gives advance notice of the exact date (Google).
• Google Cloud: dates can be extended but never moved earlier, and models in its shorter availability group get a fixed date at least 45 days out (Google Cloud).
• Microsoft Foundry: at least 60 days for generally available models and 30 days for preview models, by email to subscription owners with active deployments and as Service Health advisories (Microsoft).
• Amazon Bedrock: models launched before September 7, 2026 stay in Legacy for at least 6 months before end of life (AWS). For later models the model card shows a Legacy period of 6 months or 45 days (AWS).

Emails go to whoever owns the account or subscription, so make sure that's a shared inbox someone reads. We found no RSS feed on any of these pages. Azure comes closest: a Service Health alert rule can send an email, a text message or a webhook (Microsoft).

Why the platform you call changes the date

The same model can retire on different days depending on who serves it. Amazon Bedrock says its lifecycle dates may differ from the model provider's, and only Bedrock's dates apply on Bedrock (AWS). The Claude deprecations page says the same of Bedrock and of Google-managed Claude on Google Cloud (Claude).

• Regions can differ too. On Bedrock, anthropic.claude-3-haiku-20240307-v1:0 reaches end of life on November 13, 2026 in some Regions and November 27, 2026 in ca-central-1 (AWS).
• Azure may upgrade you. At retirement, Global Standard and Data Zone Standard deployments of gpt-4o 2024-05-13 move to gpt-5.1 automatically, and Provisioned deployments don't (Microsoft). A deployment set to NoAutoUpgrade stops working at retirement (Microsoft).
• A retired model behaves differently per platform. Requests to a retired Claude model fail (Claude), Azure returns 410 Gone (Microsoft), Bedrock requests fail after end of life (AWS), and the Gemini API routes some retired models to a newer one (Google).

So one date for a model is often the wrong date for your code. Plan against the date for the platform your code calls, and on Bedrock, the Region too.

Check the dates with a script

Three platforms put lifecycle dates in an API, so a scheduled CI job can read them. OpenAI's List models returns an optional shutdown_date for each model (OpenAI):

curl -s https://api.openai.com/v1/models \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  | jq '.data[] | select(.shutdown_date)
      | {id, shutdown_date}'

Bedrock's ListFoundationModels returns modelLifecycle, with a status of ACTIVE or LEGACY and an end of life time (AWS). Run it in each Region you call:

aws bedrock list-foundation-models \
  | jq '.modelSummaries[]
      | select(.modelLifecycle.status == "LEGACY")
      | {modelId, eol: .modelLifecycle.endOfLifeTime}'

Azure's Models API lists lifecycleStatus and deprecation.inference for every model in a region. Watch the naming: in the API, Deprecating means what the docs call deprecated, and Deprecated means retired (Microsoft).

The Claude and Gemini deprecation pages don't point to a field like these, so for those you read the page. And a script only gives you dates. It doesn't know which of your files call the model.

Our calendar API returns the dates for every provider and platform we track, free and without a key (API docs):

curl -s https://api.since.dev/v1/calendar \
  -G -d kind=model -d within=90d

Find every model ID in your code

One case-insensitive pattern catches most model IDs: OpenAI names such as gpt- and o3-, claude-, gemini-, and Bedrock IDs that start with anthropic., with or without a prefix such as us.. Add the other makers you call on Bedrock.

git grep -n -i -E \
  '(gpt-|o[134]-|claude-|gemini-|anthropic\.)'

Expect some noise, since o1- can turn up in unrelated names. Then look for IDs read from config, by the key that holds them:

git grep -n -i -E \
  "model(_?id|_?name)?[\"']?[[:space:]]*[:=]"

Then check the places a repository search misses:

• git grep reads tracked files only. Check .env files, CI secrets and your host's environment settings by hand.
• On Azure, your code sends a deployment name, which can be anything. The openai-python example passes model="deployment-name" (openai-python), so check each deployment's model and version in Foundry.
• Infrastructure code, such as Terraform stacks, and IAM policies that scope Bedrock access to a model ARN.
• Eval sets, test fixtures, recorded responses and notebooks that pin a model.
• Calls from code you don't own. The Claude Console's Usage page exports a CSV of usage by API key and model (Claude deprecations page).

Your options, compared

Where Since.dev fits

Since.dev tracks every model's retirement on each platform that serves it, from the publishers' own pages: the Claude API, OpenAI, the Gemini API, Google Cloud, Microsoft Foundry, Amazon Bedrock and Mistral. It matches each date to the platform your code calls. The retirement calendar lists the dates for free, and https://api.since.dev/v1/calendar serves them as JSON with no key, with .ics and .rss feeds.

• Scan any public GitHub repository, no sign-in needed. For a repository you scan or connect, it finds the exact files and lines that name a retiring model, with the days left and the replacement.
• A coding agent can ask through the MCP server: call check_code with a model ID, or whats_changed for what changed.
• On Free you get those impacts only. On Team and up, we open a fix pull request that swaps the model, built and checked before it opens (plans).

Fix pull requests open ready for review, your CI runs, and people decide whether to merge. Nothing merges or deploys.

Questions

How much notice do providers give before retiring a model?

It depends on the platform. OpenAI gives at least 6 months for generally available models, the Claude API at least 60 days for publicly released ones, and Azure at least 60 days for generally available models and 30 for preview. On Bedrock, models launched before September 7, 2026 stay in Legacy for at least 6 months. Each figure is from that platform's own page, linked above.

What happens if my code calls a retired model?

On the Claude API and Amazon Bedrock, the request fails. On Azure, inference returns 410 Gone, unless a Standard deployment was upgraded to the replacement. The Gemini API routes some models to a newer one at retirement, such as gemini-2.5-pro from December 1, 2026, so your requests keep working on a different model.

Is there an RSS feed for model retirements?

We found none on the OpenAI, Claude, Gemini API, Google Cloud, Microsoft Foundry or Amazon Bedrock pages. Our calendar has one, free and without a key, at https://api.since.dev/v1/calendar.rss, and an iCal feed beside it.

Do Azure, Bedrock and Google Cloud follow the provider's dates?

Not necessarily, since each sets its own. gpt-4o version 2024-08-06 leaves OpenAI's API on January 14, 2027 and runs on Azure until April 14, 2027. claude-sonnet-4-5-20250929 retires on the Claude API on November 30, 2026 and runs on Bedrock until April 8, 2027.

Do Dependabot and other dependency updaters warn about model retirements?

No. They update package versions, and a model ID is a string in your code. Our Dependabot alternative guide covers what they do and don't catch.

Related guides