Best Ways to Track LLM Deprecations and Model Retirements (2026)
The fastest way to track LLM deprecations in 2026 is to combine one primary source (the provider's own deprecation page) with one aggregator that watches all providers for you. No single provider page shows you the whole picture: OpenAI, Anthropic, Google and AWS each publish deprecations separately, in different formats, with different notice periods. Below is an honest comparison of the five approaches we know of, including what each one misses. We run an aggregator ourselves (the AI Stack Radar deprecation calendar), so read our take with that in mind; the comparison table plays it straight.
What happens when you miss an LLM deprecation?
Your production calls start failing, or get silently rerouted to a replacement model with different behavior and pricing. Our deprecation calendar currently tracks 70 dated rows across six providers (OpenAI, Anthropic, Google, AWS Bedrock, Azure Foundry and Mistral), and 39 of them fall within the next 90 days. Migration is rarely hard; being surprised is what hurts - re-testing prompts under deadline, price changes hidden in the swap, and embeddings that need re-indexing.
The five ways to track LLM deprecations, compared
| Method | Coverage | Freshness | Cost | Best for | Weakness |
|---|---|---|---|---|---|
| Provider deprecation pages OpenAI, Anthropic, Google, AWS |
Only their own models | Authoritative, first to know | Free | Verifying details before migration | Four or more pages to watch, no cross-provider view, formats differ |
| Community trackers llm-stats.com and similar |
Broad model catalogs | Good for releases, weaker for retirement dates | Free | Comparing current models | Optimized for what is new, not what dies when |
| Paid alert services e.g. DepWatch, from ~$10/mo |
Dependency and API deprecations, incl. some LLMs | Push alerts | Paid | Teams that want email or Slack pings | Cost; generic dependency focus, not LLM ops specifics |
| AI Stack Radar calendar ours, free |
OpenAI, Anthropic, Google, AWS Bedrock, Azure Foundry, Mistral in one dated table | Updated weekly with the digest | Free, no signup | Ops view: what to migrate and when | A page and RSS, not a push alerting service |
| DIY RSS + changelogs + newsletters |
Whatever you wire up | As good as your setup | Your time | Engineers who already live in RSS | Maintenance burden; our weekly digest exists because we got tired of doing this by hand |
Sources for each method: the provider deprecation pages (OpenAI, Anthropic, Google Gemini, AWS Bedrock), llm-stats.com, DepWatch, and our own deprecation calendar. Links verified July 2026.
Which method should you actually use?
It depends on who is watching and what breaks if you miss a date. Three common cases:
- Solo dev or small team: bookmark one aggregator (ours or llm-stats) and check the provider page of the one model family you depend on before each quarter. Low effort, covers the case that actually hits you.
- Ops or platform team: use paid alerts or a weekly calendar review, and put migration dates in your sprint calendar the day they are announced. The failure you are guarding against is a date nobody owned.
- Enterprise with compliance needs: provider pages are the source of record; aggregators are your early-warning layer, not your audit trail. Use both, and keep the official page as the citation.
FAQ
How much notice do providers give before retiring a model?
It varies widely. Some retirements ship with twelve or more months of notice; others, especially preview or experimental models, get only weeks. There is no industry-wide standard, so notice from one vendor tells you nothing about the next. That spread is exactly why a dated calendar beats memory: plan as if you have the short end and treat more as a bonus.
Is there a free calendar of LLM deprecation dates?
Yes. Our AI Stack Radar deprecation calendar lists dated retirements for OpenAI, Anthropic, Google, AWS Bedrock, Azure Foundry and Mistral, each with a recommended migration action per row, free and without signup. It currently tracks 70 dated rows across those providers, 39 of which fall within the next 90 days.
What is the difference between a deprecation and a retirement?
Deprecation is the announcement: the model still works and the clock starts. Retirement, or shutdown, is the date the calls stop working or get rerouted to a replacement. Plan your migration against the retirement date, and budget against the deprecation date, since a replacement can carry different pricing.
Related reading: the migration playbook for when your model is being retired, and how to keep up with AI model changes without living in changelogs.