Model Lifecycle¶
Every model in Foundry eventually retires. This page explains, in plain terms, what that means for your deployments — what happens on the day, what stays the same, and what you need to do.
GA models get about 18 months — 12 for Anthropic, DeepSeek, Fireworks and Mistral AI. The date is published on day one and can't be extended.
Global, Data Zone and regional Standard deployments are switched to the replacement in place: same name, endpoint, location and quota.
Provisioned and Batch deployments must be moved by you. If they aren't, every request fails with 410 Gone.
The five stages¶
- Preview8 Try it out May change or disappear · not for production
- Generally available48 Build on it Stable · end date known from day one
- Legacy Something better exists Still works · start looking at newer models
- Deprecated47 Existing users only Nobody new can start · plan your move
- Retired5 Switched off Every request fails with 410 Gone
Counts group tracked models by their most urgent active version; Deprecated also includes versions with a retirement due. Legacy is optional, so only a few models carry it.
- LaunchEnd date published on day one
- 12 mo · DeprecatedClosed to new users
- 18 mo · RetiredSwitched off · 410 Gone
What happens on retirement day¶
It depends on how you deployed the model. Pick your deployment type:
- Model
- o1 2024-12-17
- Type
- Data Zone Standard
- Location
- East US 2
- Endpoint
/deployments/chat-prod
- Model
- gpt-5.6-sol
- Type
- Data Zone Standard
- Location
- East US 2
- Endpoint
/deployments/chat-prod
Stays the same
- Deployment name and endpoint — your app calls the same URL
- Region or data zone — requests are processed where they were before
- Deployment type and quota — your tokens-per-minute limit carries over
Can change — test first
- The answers — it's a different model, so prompts can behave differently
- Price — tokens are billed at the new model's rate
- API features — check that the parameters and tools you use are supported
- Speed and token usage per request
Your deployment's upgrade setting decides when the swap happens
OnceNewDefaultVersionAvailableSwitches within two weeks of a new default version.OnceCurrentVersionExpiredKeeps the old model until retirement day, then switches.NoAutoUpgradeKeeps the old model — and stops working on retirement day.Find the setting under the deployment's details in the Foundry portal, or read versionUpgradeOption from the REST API, PowerShell or az cognitiveservices account deployment show. You can change it with REST, PowerShell or the portal — not with the Azure CLI. Priority Processing deployments follow the same rules.
- Model
- o1 2024-12-17
- Type
- Data Zone Provisioned
- Location
- Sweden Central
- Capacity
- 100 PTU
- Endpoint
/deployments/ptu-prod
- Model
- o1 2024-12-17
- Type
- Data Zone Provisioned
- Location
- Sweden Central
- Endpoint
/deployments/ptu-prod
- Model
- gpt-5.6-sol
- Type
- Data Zone Provisioned
- Location
- Sweden Central
- Endpoint
/deployments/ptu-prod
In place
Change the model version on the existing deployment. Azure moves traffic over a 20–30 minute window with no downtime.
Side by side
Create a new deployment on the replacement, test it, shift traffic, then delete the old one. Safest when you want a rollback path.
- The replacement becomes available in the provisioned regions where the old model is retiring about 30 days before retirement.
- PTU quota isn't tied to one model, but each model delivers different throughput per PTU — re-size with the PTU guide and make sure you have quota for the target model.
- Model
- o1 2024-12-17
- Type
- Global Batch
- Location
- East US
- Endpoint
/deployments/batch-nightly
- Deploy the replacement as a new batch deployment
- Resubmit your jobs against the new deployment
- Delete the old deployment once jobs succeed
- LaunchNot-sooner-than date set
- Notice≥30 days warning
- Upgrade or retireNewer version or ends
Upgraded to a newer preview
Your deployment is moved to the next preview version. This can repeat until a GA version exists.
Upgraded to GA
When the GA model launches, preview deployments are moved to it and follow the GA lifecycle from then on.
Removed (rare)
If there's no replacement, the model is switched off and requests return 410 Gone.
Every outcome comes with at least 30 days notice. There's no option to stay on a retiring preview model, so keep previews out of critical production.
- 1Training retires
You can no longer start new fine-tuning jobs on the base model. Models you already trained can still be deployed. This happens no earlier than the base model's retirement.
- 2Deployment retires
Inference and new deployments of the fine-tuned model return errors. Re-train on a newer base model before this date.
Fine-tuned models have their own dates — see the fine-tuned section of the retirement schedule.
Will the replacement run in my region?¶
For pay-as-you-go deployments, yes. Upgrades keep the same deployment type in the same place — and if the replacement isn't offered there yet, Microsoft adds it as part of the upgrade.
Stays global — requests can be processed in any Azure region, before and after the upgrade.
Stays in your data zone (for example the EU or the US). Same deployment type, so the same boundary applies.
Stays in the same region. If the replacement isn't offered there yet, the upgrade adds it.
Where location matters is testing early. New models arrive in Global Standard first, then Global Provisioned, then Data Zone, and regional deployments last. If your data must stay in a region or data zone, test with non-sensitive data in Global Standard — or wait until the replacement is offered where you run. Provisioned customers need the replacement to be offered in their region before they can migrate.
Is the replacement already where the old model runs?¶
For each upcoming retirement with a named replacement: of the regions where the retiring model runs today, how many already offer the replacement with the same deployment type.
- Global StandardUpgraded for you 25/25 Already offered in all 25 regions
- Data Zone StandardUpgraded for you 14/14 Already offered in all 14 regions
- Regional StandardUpgraded for you 0/7 Auto-upgrade adds it in East US, East US 2, North Central US +4 more
- Global ProvisionedYou migrate 14/29 Not offered yet in Brazil South, Canada Central, East US +12 more
- Data Zone ProvisionedYou migrate 12/12 Already offered in all 12 regions
- Regional ProvisionedYou migrate 2/4 Not offered yet in Japan East, West US 3
- Global StandardUpgraded for you 26/26 Already offered in all 26 regions
- Data Zone StandardUpgraded for you 14/14 Already offered in all 14 regions
- Global ProvisionedYou migrate 14/29 Not offered yet in Brazil South, Canada Central, East US +12 more
- Data Zone ProvisionedYou migrate 13/13 Already offered in all 13 regions
- Regional ProvisionedYou migrate 2/8 Not offered yet in East US, East US 2, North Central US +3 more
- BatchYou migrate 0/22 Not offered yet in Australia East, Brazil South, Canada East +19 more
- Global StandardUpgraded for you 2/2 Already offered in all 2 regions
- Global StandardUpgraded for you 25/25 Already offered in all 25 regions
- Data Zone StandardUpgraded for you 16/16 Already offered in all 16 regions
- Global ProvisionedYou migrate 28/29 Not offered yet in North Europe
- Data Zone ProvisionedYou migrate 14/14 Already offered in all 14 regions
- Regional ProvisionedYou migrate 0/10 Not offered yet in Australia East, Canada East, East US 2 +7 more
- BatchYou migrate 0/22 Not offered yet in Australia East, Brazil South, Canada East +19 more
- Global StandardUpgraded for you 3/3 Already offered in all 3 regions
- Global StandardUpgraded for you 26/26 Already offered in all 26 regions
- Data Zone StandardUpgraded for you 14/14 Already offered in all 14 regions
- Regional StandardUpgraded for you 0/7 Auto-upgrade adds it in East US, East US 2, North Central US +4 more
- Global ProvisionedYou migrate 28/29 Not offered yet in North Europe
- Data Zone ProvisionedYou migrate 13/13 Already offered in all 13 regions
- Regional ProvisionedYou migrate 1/10 Not offered yet in Central US, East US, East US 2 +6 more
- BatchYou migrate 0/22 Not offered yet in Australia East, Brazil South, Canada East +19 more
- Global StandardUpgraded for you 25/25 Already offered in all 25 regions
- Data Zone StandardUpgraded for you 11/16 Auto-upgrade adds it in Germany West Central, Italy North, Poland Central +2 more
- Regional StandardUpgraded for you 9/18 Auto-upgrade adds it in Australia East, Canada East, France Central +6 more
- Global ProvisionedYou migrate 29/29 Already offered in all 29 regions
- Data Zone ProvisionedYou migrate 12/14 Not offered yet in US Gov Arizona, US Gov Virginia
- Regional ProvisionedYou migrate 12/29 Not offered yet in Brazil South, Canada Central, Central US +14 more
- BatchYou migrate 22/22 Already offered in all 22 regions
- Global StandardUpgraded for you 8/8 Already offered in all 8 regions
- Data Zone StandardUpgraded for you 2/2 Already offered in all 2 regions
- Global StandardUpgraded for you 31/31 Already offered in all 31 regions
- Global StandardUpgraded for you 31/31 Already offered in all 31 regions
- Data Zone StandardUpgraded for you 0/16 Auto-upgrade adds it in Central US, East US, East US 2 +13 more
84 other models retire later without a named replacement yet — Microsoft names one 90–120 days before the date. Track them in the retirement planner.
From the latest availability snapshot. Regions are compared by model name, not by individual version. Gaps in pay-as-you-go rows are filled by the upgrade itself; gaps in provisioned or batch rows mean you can't migrate there yet.
Countdown to a retirement¶
-
−120 to −90 days
Replacement named
Microsoft picks the recommended replacement and lists it in the retirement schedule.
You: Start testing newer models — you don't have to wait for this.
-
≈ −90 days
Testable in Global Standard
The replacement can be deployed in Global Standard.
You: Run your own prompts and evaluations against it.
-
≥ −60 days
You get told
Email to subscription owners and an Azure Service Health advisory.
You: Set a Service Health alert so the right people see it.
-
≈ −30 days
PTU window opens
The replacement appears in the provisioned regions where the old model retires.
You: Migrate provisioned deployments in place or side by side.
-
Day 0
Retirement
Pay-as-you-go deployments are upgraded region by region. Anything not moved returns 410 Gone.
You: Dates can't be extended.
Sent automatically to subscription owners with active deployments.
Health advisories under Azure OpenAI Service — create an alert rule for email, SMS or a webhook.
Check lifecycleStatus and deprecation dates for any model, any time.
Common questions¶
Can I get more time before a model retires?
No. Retirement dates can't be extended. In rare cases a model with a security or compliance problem can be retired early with shorter notice.
Do I have to use the replacement Microsoft names?
No. It's Microsoft's recommendation and the model used for automatic upgrades, but you can migrate to any model that suits you. Compare quality, speed and cost on your own prompts rather than public benchmarks.
Will my code need to change?
Usually not to keep calling the deployment — the name and endpoint stay the same. But check that the API parameters, tools and output format you rely on still behave the same with the new model.
What does 'existing customer' mean for deprecated models?
It's decided per Azure subscription: a subscription that has ever deployed that exact model version can keep creating deployments until retirement. A new subscription in the same tenant doesn't inherit that access.
Is Azure Government different?
Yes. Global Standard isn't available, not every model is offered, and usually only one version of a model is available at a time, with a 30-day overlap when a new version arrives.
Reference¶
What each badge on this site means
Hover or tap any lifecycle badge on the site for this explanation in place.
| Badge | What it means | What to do |
|---|---|---|
| Retires in 12 days | Firm retirement date within 30 days | Move traffic to the replacement now |
| Retires in 60 days | Firm retirement date within 90 days | Test the replacement and plan the cut-over |
| Retirement imminent | A no-earlier-than date has passed — it can retire any time after notice | Treat as retiring now |
| Deprecated | Closed to new customers | Don't start new work on it |
| GA · until Jun 2027 | Generally available, with its next retirement month | Safe to build on |
| No retirement date | Not in Microsoft's retirement tables yet (common for new and partner models) | Check the model card in Foundry |
Who can deploy at each stage
| Stage | Meaning | New deployments | Existing deployments |
|---|---|---|---|
| Preview | Experimental — weights, runtime and API might change; not guaranteed to reach GA | Yes | Yes |
| Generally available | Production-ready — weights and APIs are fixed | Yes | Yes |
| Legacy | Newer, more capable models exist (optional stage) | Yes, until deprecated | Yes |
| Deprecated | No longer available to new customers | Only subscriptions that already used this version | Yes |
| Retired | Removed from service — every request returns 410 Gone |
No | No |
Reading the Models API (the names don't match the portal)
The API's Deprecated means retired, not deprecated — watch out for this when scripting checks.
API lifecycleStatus |
Means | Badge on this site |
|---|---|---|
Preview |
Preview, not for production | Preview |
GenerallyAvailable |
GA and open to new customers | Generally available |
Deprecating |
Deprecated — existing customers only | Deprecated |
Deprecated |
Retired — no longer served | Retired |
If deprecation.inference is in the past, treat the model as retired even if lifecycleStatus hasn't caught up yet.
Adapted from Microsoft Learn: Foundry Models lifecycle and support policy, Working with models and Model migration process. See Retirements for every model-specific date.