The calendar
From the Foundry model retirement schedule (updated September 2026). Dates can change; check the schedule before acting.
| Date | Model | Version | Replacement (if named) |
|---|---|---|---|
| 2026-11-15 | codex-mini | 2025-05-16 | — |
| 2026-11-15 | claude-haiku-4-5 | 1 and 2 | — |
| 2026-11-15 | Cohere-embed-v3-english / -multilingual | 1 | Deprecated on this date (retire 2027-02-15); Cohere-embed-v5-fast / -pro |
| 2026-11-17 | gpt-chat-latest (preview) | 2026-05-05, 2026-06-24 | Preview: force-upgraded or retired |
| 2026-11-19 | o1, o1-pro, o3, o3-pro, o3-deep-research | various | gpt-5.6-sol |
| 2026-11-19 | o3-mini, o4-mini | various | gpt-5.6-terra |
| 2026-11-19 | Llama-4-Maverick-17B-128E-Instruct-FP8 | — | — |
| 2026-11-24 | claude-opus-4-5 | 1 | — |
| 2026-11-30 | claude-sonnet-4-5 | 1 | claude-sonnet-5-5 |
| 2026-12-08 | gpt-chat-latest (preview) | 2026-05-28 | — |
| 2026-12-09 | gpt-4o | 2024-05-13 | gpt-5.6-sol (Standard auto-upgrade) |
| 2026-12-09 | Ministral-3B | 1 | ministral-ministral-4b-instruct |
| 2026-12-09 | mistral-medium-2505 | 1 | mistral-medium-3-5 |
| 2026-12-15 | whisper, tts-hd (GA), tts (preview) | 001 | — |
| 2026-12-16 | gpt-image-1.5 | 2025-12-16 | — |
Details on the o-series: o-series retirement runbook.
What each kind of retirement means
| Situation | What happens | Your move |
|---|---|---|
| Azure OpenAI GA model, Standard deployment, auto-upgrade allowed | Upgraded to the replacement at retirement | Test the replacement now; switch on your schedule |
Azure OpenAI GA model, NoAutoUpgrade or no replacement named | Requests return 410 Gone after the date | Migrate before the date |
| Provisioned deployment | Never auto-upgraded | Manual migration; check PTU quota |
Preview model (gpt-chat-latest previews) | Force-upgraded to a newer preview/GA, or retired, with at least 30 days' notice | Don't build production on previews |
| Partner models (Anthropic, Mistral, DeepSeek, Fireworks) | 12-month GA lifecycle instead of 18; Fireworks per-token models have 15-day notice | Plan model switches as a routine, not a project |
| Embedding models (Cohere embed v3) | Deprecated first, retired later | Re-embed your index - see below |
Embedding retirements are different
A chat model swap changes answers. An embedding model swap invalidates every vector you've stored: vectors from different models aren't comparable. Cohere embed v3 is deprecated on 15 November 2026 and retires on 15 February 2027, so any vector index built with it needs a full re-embed before February. See re-embedding a vector index without downtime.
Claude models on Foundry
Claude on Foundry comes in two hosting versions (1: hosted by Anthropic, 2: hosted on Azure). Where the schedule lists both versions of a model, they share the retirement date in the table. If you call Claude through Foundry:
- claude-haiku-4-5 (15 Nov) has no named replacement in the schedule; evaluate a current Haiku or Sonnet model.
- claude-sonnet-4-5 (30 Nov) lists claude-sonnet-5-5 as the replacement.
- claude-opus-4-5 (24 Nov): evaluate a current Opus model.
Check each model's parameters and tool-use behaviour on your prompts before switching.
Quick inventory
az cognitiveservices account list --query "[?kind=='OpenAI' || kind=='AIServices'].[name,resourceGroup]" -o tsv |
while read n g; do
az cognitiveservices account deployment list -n "$n" -g "$g" \
--query "[].[name, properties.model.name, properties.model.version, sku.name]" -o tsv | sed "s/^/$n\t/"
done
Match the output against the table.
Checklist
- Every deployment matched against the calendar
- Owner and migration date per affected deployment
- Replacement evaluated on your own prompts
- Vector indexes built with retiring embedding models scheduled for re-embedding
- Nothing in production on preview models