LLM providers are retiring models faster than you can migrate

On May 15, 2026, xAI retired 8 Grok API models. The notice period was 9 days.

If you had grok-2, grok-3, or grok-4-fast pinned in production, here's the part that actually bites: the retired slugs don't hard-error. They silently redirect to grok-4.3 — reasoning models drop to low effort, non-reasoning to none — and you get billed at grok-4.3 pricing ($1.25 / $2.50 per 1M tokens). xAI's original retirement email said the requests would "no longer work"; a later docs update introduced the silent-redirect behavior. The two are contradictory, and either way your output quality and your bill changed without a single error in your logs.

This is not an xAI problem. It's the whole industry right now:

• 
  OpenAI removed chatgpt-4o-latest from the API on Feb 17, 2026. The Assistants API sunsets Aug 26, 2026.


• 
  Anthropic ends Claude Opus 4 and Sonnet 4 on Jun 15, 2026 (Opus 3 already retired Jan 5; Haiku 3 on Apr 19).


• 
  Google can shut off Gemini 2.0 Flash / Flash-Lite as early as Jun 1, 2026; they've been restricted to existing customers since Mar 6.

Pinning model IDs is correct — and it's now a liability

The standard advice is to pin explicit model versions for reproducibility, and that advice is right. A floating alias means your behavior changes silently under you. But a pinned ID means that when the provider retires it, you break — or worse, get silently rerouted. Either way, the failure mode is the same: something you depend on changed, and nobody told you in a channel you actually watch. Provider changelogs are scattered across docs pages, status pages, dashboard banners, and one-off emails to whatever address owns the billing account.

What I did about it

I started keeping a normalized, cross-provider timeline of every deprecation / breaking change / pricing change I could verify against provider docs: provider, model, event type, announced date, effective date, recommended replacement, source link.

• Timeline:
  

It currently covers OpenAI, Anthropic, Gemini, and xAI. It's manually curated and fact-checked right now — the data is the hard part, not the page. If you spot a missing or wrong event, or a provider you want covered, tell me and I'll add it.

If you run anything against an LLM API in production, subscribe to the RSS feed or check it before your next deploy. The next 9-day notice is already on the calendar.