The short answer
A preview model can serve production traffic only when three things hold: the feature tolerates a forced model swap, a generally available (GA) model is wired in as a tested fallback, and the preview does something no GA model does for you. Everything else should stay off previews.
The reason is not that previews disappear quickly. Most last months. The reason is the notice you get when they do: at least 30 days on Microsoft Foundry for previews, against at least 60 days for GA models 1, and 14 days on Alibaba Cloud Model Studio, against 3 months for stable models 2. Foundry's own page says it does not recommend preview models in production 3.
Today is 18 November 2026. Several of the previews named below are inside or past their retirement dates, so the dates are worth re-checking against the live schedule.
How long previews were scheduled to last
Foundry's schedule lists a version and a retirement date for each model. Subtracting one from the other gives a rough scheduled lifetime. It is rough because the version date may precede public availability, and because Foundry says preview retirement dates are sometimes extended 1. The schedule page was last updated 2026-09-23 4. The table lists preview versions, sorted by days; day counts are calculated from the schedule's dates 4.
| Model | Version date | Scheduled retirement | Days |
|---|---|---|---|
| MAI-Image-2.5-Pro | 2026-06-19 | 2026-10-01 | 104 |
| MAI-Image-2.5 | 2026-06-02 | 2026-10-01 | 121 |
| MAI-Image-2.5-Flash | 2026-06-02 | 2026-10-01 | 121 |
| DeepSeek-V4-Flash-0731 | 2026-07-31 | 2026-12-03 | 125 |
| gpt-chat-latest | 2026-06-24 | 2026-11-17 | 146 |
| gpt-chat-latest | 2026-08-06 | 2027-01-12 | 159 |
| gpt-chat-latest | 2026-05-28 | 2026-12-08 | 194 |
| gpt-chat-latest | 2026-05-05 | 2026-11-17 | 196 |
| gpt-chat-latest | 2026-10-07 | 2027-05-18 | 223 |
| MAI-Transcribe-1 | 2026-01-23 | 2026-09-15 | 235 |
| gpt-realtime-2 | 2026-05-06 | 2027-01-20 | 259 |
| sora-2 | 2025-10-06 | 2026-07-15 | 282 |
| Kimi-K2.7-Code | 2026-06-12 | 2027-04-16 | 308 |
| sora-2 | 2025-12-08 | 2026-10-15 | 311 |
| Kimi-K2.6 | 2026-04-20 | 2027-04-16 | 361 |
| model-router | 2025-08-07 | 2027-01-27 | 538 |
| gpt-image-1 | 2025-04-15 | 2026-10-23 | 556 |
| model-router | 2025-05-19 | 2027-01-27 | 618 |
Three readings of the table:
- Nothing was scheduled under about 100 days. That matches the stated pattern: previews launch with a "not sooner than" date, typically 90 days out, and are sometimes extended until a replacement exists 1.
- The median is about 229 days, and 8 of 18 were scheduled under 200 days. Plan for a preview living one to two quarters, not a year.
- The long tail is real but not something to plan on. The model-router previews run past 500 days, and are due to retire on 27 January 2027 4; see the model-router post.
Two of the previews you may be running have no version date to measure. Cohere-parse-v5 is a preview scheduled to retire 2026-12-15 with no replacement listed. Kimi-K2.5 is a preview scheduled to retire 2027-01-26, also with no replacement listed 4. For Kimi-K2.5, the retired-models page shows it was the suggested replacement for Kimi-k2-thinking, which retired on 29 March 2026 5. A model can be a migration target and a preview at once.
What notice to expect
Notice is the number that decides whether a preview is survivable. Lifetimes are long enough to ship against, but the warning is short.
| Platform | Class | Notice before retirement |
|---|---|---|
| Foundry | GA | At least 60 days, sent to subscription owners with active deployments 1 |
| Foundry | Preview | At least 30 days 1 |
| Foundry | Fireworks per-token | 15-day notice period 4 |
| Model Studio | Stable | No later than 3 months before end of service 2 |
| Model Studio | Preview | No later than 14 days before end of service 2 |
Other facts from the same pages that change how you plan:
- Foundry reserves the right to run an emergency retirement with shortened notice for compliance or security problems 1.
- After retirement, Foundry inference returns 410 Gone 1. Your code needs a path for that status, not just for 5xx.
- On Model Studio, rate limits for a retiring model may be reduced gradually from the start of retirement, before the service stops 2. You can see throttling before the shutdown date.
- Model Studio also says a retirement may be paused, postponed or cancelled 2. Do not treat a date as final in either direction.
- On Model Studio, a preview is identified by preview, beta or experimental in the ID, or a "Preview" label. A
-latestalias counts as a stable model there 2. On Foundry,gpt-chat-latestis a Preview 3. A "latest" suffix tells you nothing about lifecycle class.
Three traps in the replacement column
The suggested replacement can be another preview. Foundry's retired chat models (gpt-5-chat, gpt-5.1-chat, gpt-5.2-chat, gpt-5.3-chat) all list gpt-chat-latest as the suggested replacement 5, and gpt-chat-latest is a Preview with five versions live at once 3. In the same schedule, the GA DeepSeek-V4-Flash (retiring 2028-02-20) lists DeepSeek-V4-Flash-0731 as its replacement, and that is a preview scheduled to retire 2026-12-03 4. Following the replacement column mechanically can move you from a stable line to a short-lived one.
A forced upgrade is not a choice. When a preview retires, deployments are force-upgraded to a newer preview or to GA, or terminated if there is no replacement. You cannot remain on the old version 1. The weights, runtime and API schema of a preview can change, and it is not guaranteed to become GA 1. Your outputs can change on a date you did not pick.
A replacement can mean a different hosting model. The Cohere-command-a-plus-05-2026 preview (scheduled 2026-10-16) lists replacements that are Hugging Face collection models for managed compute deployment 4. That is an infrastructure change, not a model-id swap.
A policy: where previews are allowed
Thresholds below are policy suggestions, not vendor facts. Adjust them to your own cost of a bad day.
| Tier | Where | Rule |
|---|---|---|
| 0: not allowed | Money movement, legal or medical output, anything that triggers an action without review, outputs parsed by strict downstream schemas | No preview as any target, including fallback |
| 1: guarded | Customer-facing features with a human in the loop or easy retry | Preview allowed only behind a GA fallback with a tested equivalence check; owner and exit date recorded |
| 2: open | Internal tools, prototypes, evals, staging | Preview allowed; no fallback needed; re-checked each time the schedule changes |
Three rules apply to every tier:
- Record the retirement date and the notice floor at approval time. A preview with no date in your register is a preview you will find out about by email.
- Set an internal cut-over date well inside the notice floor. With a 30-day floor, a 45 to 60 day lead means the swap is a planned change, not an incident.
- Re-approve on every force-upgrade. The model answering after an upgrade is a different model. Treat it as a new approval.
Routing guardrails
Policy only holds if the request path enforces it. The guardrails that matter:
- Keep previews out of production by default. Scope preview routes to staging and development, and let production keys reach only GA targets unless a Tier 1 approval exists.
- Allowlist by project. A project that has no Tier 1 approval should get a policy error when it names a preview, not a line item.
- Never make a preview the only target. Put the GA model second at minimum. Make the failure rules cover a retired model's 410 as "move on", not "stop".
- Send only eligible traffic to the preview. Match on feature or metadata so the preview sees the tolerant slice, not everything.
- Watch for the silent swap. A force-upgrade produces no error. Detect a change in the answering model, longer answers or a new error mix.
- Check availability from your own key. A model your key stops listing is the earliest hard signal that it is gone.
A small CI check keeps the register honest. This sketch is illustrative; load your own schedule export.
from datetime import date
# Illustrative register: model id -> (tier, retirement date, internal cut-over lead in days)
REGISTER = {
"example-preview-a": (1, date(2026, 12, 15), 45),
}
def check(model: str, today: date) -> str:
tier, retires, lead = REGISTER.get(model, (None, None, None))
if tier is None:
return "block: preview not in register"
days_left = (retires - today).days
if days_left <= 0:
return "block: past scheduled retirement"
if days_left <= lead:
return f"warn: {days_left} days left, inside the {lead}-day cut-over lead"
return "ok"
print(check("example-preview-a", date(2026, 11, 18)))Enforcing it on your own provider connections
Several of these guardrails are shipped INFRO capabilities that run over the provider accounts you already hold. Environments keep a route or connection to production, staging or development, so a preview route can sit in staging only. Model allowlists return a policy error for an unapproved model. Routing rules send matching requests by feature or metadata to their own targets, and failure rules decide per failure kind whether to retry, move on or stop. Model lifecycle records vendor deprecations and shutdown dates, warns organizations still sending a retiring model, and from the shutdown date refuses it before any provider call, naming the replacement. Model availability reads the list your own key returns every three hours and records the date a model stopped appearing. Traffic changes flags a model answered by a new snapshot. INFRO sees request metadata, not answer content, so it does not judge whether the new answers are worse. See the INFRO docs (opens in a new tab) for the request contract.
Exit checklist for previews you already run
- List every preview model id in production traffic, including aliases ending in
latest. - Record each one's scheduled retirement date and the vendor's notice floor.
- For
gpt-chat-latestversions dated 2026-05-05 and 2026-06-24, confirm what answered after their 2026-11-17 date 4. - For DeepSeek-V4-Flash-0731 (2026-12-03), Cohere-parse-v5 (2026-12-15) and Kimi-K2.5 (2027-01-26), name the GA or stable target and run it against a saved set of real inputs 4.
- Make sure no replacement you picked is itself a preview.
- Add a failure rule for 410 Gone that moves to the fallback.
- Put the cut-over date in a calendar, not in an inbox.
For the one-day-notice case, see the fallback plan; for alias-versus-snapshot trade-offs, see aliases vs pinned snapshots.
