What the date does to a deployment
Azure's lifecycle has five stages, and the last one is blunt: once a model is Retired, every inference request returns 410 Gone, and no new deployments can be created.1 That is the failure you can alert on. The other outcome produces no error at all. For Global Standard, Data Zone Standard and Standard deployments, Microsoft manages automatic upgrades when a version retires, rolling out region by region.1 The deployment name stays the same and the model behind it changes.
Which of the two you get depends on three things you can read today: the deployment type, the version upgrade option, and whether the schedule lists a replacement. Dates alone do not tell you that. This checklist starts from the behaviour and derives the routing rule, then applies it to the three dates in play.
Today is 2026-11-05. The o-series date is 14 days away, whisper and tts are 40 days away, and GPT-5 is 96 days away.
The three dates and what Microsoft lists
The schedule page, last updated 2026-09-23 when read, gives these rows.2
| Deployed model | Retires | Listed replacement |
|---|---|---|
| o1, o1-pro, o3, o3-pro, o3-deep-research | 2026-11-19 | gpt-5.6-sol |
| o3-mini, o4-mini | 2026-11-19 | gpt-5.6-terra |
| whisper 001, tts 001, tts-hd 001 | 2026-12-15 | none listed |
| gpt-5 and gpt-5-nano, 2025-08-07 | 2027-02-09 | none listed |
Microsoft says the replacement column is its recommended target and, where supported, the target for automatic upgrades.2 A dash therefore means you cannot read an upgrade target off the page. Microsoft also says it names the official replacement roughly 90 to 120 days before retirement, not sooner.1 For Dec 15 that window has already passed with nothing listed. For Feb 9 it runs from about Oct 12 to Nov 11, so a row could appear any day. Re-read the page before you finalise the GPT-5 rule.
Four deployment states, four routing rules
The row to find first is your deployment's own state. The upgrade option has three values: OnceNewDefaultVersionAvailable, OnceCurrentVersionExpired and NoAutoUpgrade. A missing value is null and behaves as OnceCurrentVersionExpired.3 Under NoAutoUpgrade the deployment stops working at the retirement date.3
| State on the date | What Azure does | What your client sees | Routing rule |
|---|---|---|---|
Standard-type, null or OnceCurrentVersionExpired | Upgrades to the default version at retirement3, region by region1 | Success, different model | Canary the replacement before the date; watch for a changed model, not for errors |
Standard-type, NoAutoUpgrade | Stops working3 | 410 Gone1 | Explicit fallback target on a different deployment |
| Provisioned | Not auto-upgraded; you migrate manually1 | 410 Gone after the date | Replacement deployment plus a Standard overflow target |
| Preview | Force-upgraded or terminated, at least 30 days notice1 | Swap or 410 Gone | Do not route production traffic here |
I could not confirm from the pages read how the Standard-type auto-upgrade and a NoAutoUpgrade setting interact beyond the sentence quoted above. Test it on one non-production deployment rather than assuming.
Why the silent swap is the dangerous one
A fallback rule fires on failure. An auto-upgrade is a success with a different model: different reasoning behaviour, token usage and output shape, and nothing to retry. Microsoft notes upgrades can happen even if the new version is not yet separately available in that region.1 Because they roll region by region, two regional deployments behind one route can answer with different models for a while. If you want determinism, set NoAutoUpgrade on deployments you have not evaluated, and accept that you then own the failure path.
Checklist: before the date
Run these per deployment, not per model.
- List every deployment of the retiring models, with its deployment type (Global Standard, Data Zone Standard, Standard or provisioned), region and subscription.
- Read each deployment's
versionUpgradeOption. The Azure CLI can read it but cannot update it; use REST, PowerShell or the portal to change it.3 - Query lifecycle fields from the Models API (
lifecycleStatus,deprecation.inference, per-SKUdeprecationDate) and diff them against your own registry nightly.1 - Decide per deployment: move it yourself to an evaluated replacement, let Azure upgrade it, or set
NoAutoUpgradeand fail over. - Create the replacement deployment now, in each region and deployment type that serves traffic, and send it a small share of real traffic.
- Confirm API compatibility, feature support, regional availability, capacity and quota for the target, as Microsoft advises.1
- Forward retirement notices to the route's owner: Microsoft emails subscription owners, which may be a different team.1
az cognitiveservices account deployment show \
--resource-group "$RG" --name "$ACCOUNT" --deployment-name "$DEPLOYMENT" \
--query "properties.{model:model.name,version:model.version,upgrade:versionUpgradeOption}"A null upgrade value means the option was never set, which behaves as OnceCurrentVersionExpired.3
Quota: the replacement is a different pool
Since May 7, 2026 Foundry tracks quota at the subscription level. Global Standard deployments of the same model and version share one pool across all regions; Data Zone Standard shares one per data zone.4 By that rule a deployment moved to a different model draws on the replacement's pool, not the o-series pool it came from. I could not confirm how capacity is carried across an auto-upgrade, so check the assigned capacity after the move. A capacity value of 1 equals 1,000 tokens per minute.3
Tier 1 Global Standard limits show why this matters:4
| Model | RPM | TPM |
|---|---|---|
| o3 | 1,000 | 1,000,000 |
| o4-mini | 1,000 | 1,000,000 |
| o3-mini | 500 | 5,000,000 |
| gpt-5.6-sol | 250 | 250,000 |
| gpt-5.6-terra | 250 | 250,000 |
Quota tiers rise automatically with usage, and your subscription may sit higher than Tier 1.4 Look up your tier through the control plane API before assuming. If the replacement's limits are lower than your peak, an upgrade turns into 429 responses on the day, which is exactly the case a fallback target should absorb.
Failover rules that fit each behaviour
Keep the application calling a stable name rather than a deployment id. In INFRO a route is a name your application sends as model, resolving to an ordered list of your own connections and the model each should serve.5 Priority routes walk targets in order; weighted routes choose the first target in proportion to weight and still fail over, and a weight-0 target is a standby.5
An illustrative layout, not INFRO's configuration syntax:
route: support-reasoning
phase_1_before_nov_19:
primary: azure-o3-deployment # retiring
canary: azure-gpt-5-6-sol-deployment # small weight, measured
phase_2_after_cutover:
primary: azure-gpt-5-6-sol-deployment
standby: second-region-or-provider # different failure domain
rule: on 429 or 5xx before first token -> next targetThree points follow from the behaviours above.
- Put the canary before the date. It is the only control that touches the silent-swap case.
- Make the standby a different failure domain. A second deployment of the same retiring model fails with the first. Connection failover is same-model, and model fallbacks are a separate, opt-in list of up to five catalog models.6
- Do not assume
410triggers failover. INFRO's docs list 5xx, timeout, 429, 401/403, 404 and 400/422 as upstream failures that move on to the next connection.6 A410is not named on the page I read. Test it in a drill, and if it does not move on, add a failure rule for that kind; failure rules decide per failure kind whether to retry, move on or stop.
For whisper, tts and tts-hd, the fallback must be an audio model that takes the same request shape. A text model cannot stand in, so verify the standby accepts your voice, format and language parameters before the date. The same goes for GPT-5 with no listed replacement: pick the target by evaluating it on your prompts, as Microsoft advises.1
Verify after the date
- Count requests by status for each deployment for 24 hours: any
410means a deployment you missed. - Compare the model and version each deployment reports against the one you expected, in every region.
- Check latency, output length and cost per request against the week before; an upgrade shifts all three without an error.
- Confirm the standby took traffic in a forced test, not only on paper.
Applying this with a gateway
INFRO's model retirements capability records each vendor's deprecations, warns organizations still sending a retiring model, and refuses it from the shutdown date before any provider call, naming the replacement. Those are vendor dates, and Azure keeps its own, so keep the Azure rows above in your own registry too. Routes, failover and routing rules apply the layouts in this article over your own Azure connection. Traffic changes flags a model answered by a new snapshot and longer answers, with an estimated weekly cost change; it sees request metadata, not answers, so it does not judge quality. The INFRO docs overview (opens in a new tab) describes the control plane;7 request failures that reach you use one error shape.8
