The dated list
Microsoft's retirement schedule marks all of these as Preview and lists no replacement for any of them.1 The first date is three days away and falls on a Tuesday, so treat Monday as your real deadline.
| Retires | Model | Stage | Listed replacement |
|---|---|---|---|
| 2026-11-24 | MAI-Image-2.5, MAI-Image-2.5-Flash, MAI-Image-2.5-Pro | Preview | none |
| 2026-12-04 | MAI-Transcribe-2-Streaming | Preview | none |
| 2027-02-04 | Microsoft-Decision-1 | Preview | none |
Once a model retires, all inference requests return 410 Gone and no deployments can be created.2 Preview deployments are never left in place: they are force-upgraded to a newer preview or GA version, or terminated if nothing replaces them, with at least 30 days' notice.2 With no replacement listed, assume termination.
The schedule's own guidance applies: you don't have to wait for Microsoft to name a replacement, and you should evaluate candidates on your own application and data.1 The targets below are candidates drawn from the same pages, not Microsoft recommendations.
Image: MAI-Image-2.5 family, Nov 24
The nearest GA entries in the family are MAI-Image-2.6 and MAI-Image-2.6-Flash, both GA with a listed retirement of 2027-05-01.1 That is about five months away, so this is a bridge. Schedule a second review before then. No GA "Pro" tier appears in the schedule, so if you call MAI-Image-2.5-Pro, check whether 2.6 meets your quality bar. Don't assume it does.
Prompts that were tuned on 2.5 may render differently on 2.6. Check size and aspect-ratio support, content-filter behaviour and output format, since any of these can fail without a model-name error.
Smoke test: send five to ten of your real production prompts, at the sizes you actually request, to the candidate. Pass means an HTTP 200, a decodable image, the requested dimensions, and no moderation refusal on prompts that passed before. Do this before you edit any route.
Streaming speech: MAI-Transcribe-2-Streaming, Dec 4
This one has the later date but the least obvious target. The schedule lists no row for MAI-Transcribe-1.5 beyond naming it as the replacement for MAI-Transcribe-1, so confirm its lifecycle before you commit to it. GA transcription models on the schedule include gpt-live-transcribe and gpt-transcribe (both listed for 2028-02-01) and gpt-realtime-whisper (2027-11-09).1 I did not confirm which of these accept streaming input, so read each model's page before testing.
If your gateway sits in front of transcription, check its shape. INFRO's transcription endpoint takes an uploaded recording and is synchronous, so it does not carry a live stream.3 A streaming client has to keep its direct connection to the speech endpoint, and that connection needs the new model name.
Smoke test: replay a recorded call of known length through the candidate as a live stream. Pass means partial results arrive while audio is still being sent, the final transcript's word error rate on your reference set is within your threshold, and the session closes cleanly on both normal end and a dropped socket.
Decisions: Microsoft-Decision-1, Feb 4
You have the most time here, which makes it the easiest to forget. Microsoft-Decision-1 is one of the decision models INFRO can run on your own Azure connection, mapped to your Foundry deployment name. The same docs list other decision models on other providers, including GPT-6 Luna (Decisions), Clef, pplx-decider, Mercury Decide and Solar Decide.4 A move to any of them means a different provider account, so check its data handling and contract terms before testing.
Decisions are typed outputs such as a probability or one option from a closed set. The risk is calibration. A confidence of 0.8 from one model is not 0.8 from another, so thresholds set against Decision-1 need re-fitting.
Smoke test: replay a labelled set of past decisions through the candidate. Pass means every output parses into your schema, accuracy against your labels meets the bar, and the share of items falling in your "unsure" band is close to what it was on Decision-1.
A harness that fails loudly
Keep the check provider-neutral so you can run it against both old and new targets. The part that matters here is treating 410 as retirement, not a transient error.
import os, sys, requests
URL = os.environ["TARGET_URL"] # your candidate endpoint
HEADERS = {"Authorization": f"Bearer {os.environ['TARGET_KEY']}"}
CASES = [...] # your recorded production requests
failed = 0
for i, body in enumerate(CASES):
r = requests.post(URL, json=body, headers=HEADERS, timeout=60)
if r.status_code == 410:
print(f"case {i}: 410 Gone, target is retired"); failed += 1
elif r.status_code != 200:
print(f"case {i}: HTTP {r.status_code}"); failed += 1
# add per-workload assertions: image size, schema parse, WER
sys.exit(1 if failed else 0)Where a gateway helps
INFRO records vendor retirement dates and, from the shutdown date, refuses a retired model before any provider call and names the replacement. That turns a surprise 410 from your provider into a clear error you can alert on. It is no substitute for the smoke tests above, and it only knows models it lists. Check your own model list against the schedule today.
