Azure OpenAI GPT-5.5 retirement: GPT-5.4 on 2 Sept, GPT-5.5 on 26 Oct 2027

Summary. Microsoft's model retirement schedule, last updated on 23 September 2026, retires gpt-5.4 (version 2026-03-05) on Azure OpenAI on 2 September 2027, gpt-5.4-pro on 7 September, gpt-5.4-mini and gpt-5.4-nano (version 2026-03-17) on 21 September, and gpt-5.5 (version 2026-04-24) on 26 October 2027. After that, requests return 410 Gone, and the dates cannot be extended. A quieter deadline comes first: the lifecycle policy closes a GA model to new subscriptions 12 months after launch, about March 2027 for the GPT-5.4 family and April 2027 for gpt-5.5. No replacement is named yet; Microsoft names one 90 to 120 days before retirement. Standard deployment types auto-upgrade unless set to NoAutoUpgrade; provisioned and Batch deployments never do. At Azure Global Standard rates, gpt-5.4 costs $2.50 input and $15 output per million tokens and gpt-5.5 costs $5 and $30, against $2 and $12 for gpt-5.6-terra and $4 and $20 for gpt-5.6-sol, both kept until 11 January 2028. OpenAI calls that Sol rate promotional until 30 November 2026. OpenAI's own API drops gpt-5.4-nano on 1 April 2027. Inventory now, lock upgrade settings and move tool calls to the Responses API before testing candidates.

About this article. For IT admins, platform teams and engineering leads running GPT-5.4 or GPT-5.5 on Azure OpenAI in Microsoft Foundry. On 7 November 2026 we checked dates, lifecycle rules and regional availability against Microsoft Learn's retirement schedule, lifecycle policy, model and region pages, every price against the Azure Retail Prices API, and OpenAI's deprecations and pricing pages. We did not query the Models API from a live subscription. Schedules and prices change often, so re-check before you commit; this is not procurement advice, and no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.

What retires on Azure, and when

The schedule gives one retirement date per model version. The table adds how long each version lives, the 12-month new-customer cutoff implied by Microsoft's lifecycle policy, and OpenAI's status.

The cutoff column is our arithmetic: the policy moves a model to Deprecated 12 months after launch, but Microsoft prints no per-model cutoff date, and stage labels lag. The same schedule still shows gpt-5 (2025-08-07) as GA more than a year after launch. Treat the cutoff as a planning date and verify it from a fresh subscription.

The OpenAI column matters for multi-cloud setups. OpenAI's 1 October 2026 notice removes gpt-5.4-nano and gpt-5.4-nano-2026-03-17 from its API on 1 April 2027 and recommends gpt-6-luna; our guide to the April 2027 OpenAI API shutdowns covers the rest of that notice. Azure keeps gpt-5.4-nano 173 days longer, so a cross-cloud fallback for nano disappears in April.

Older versions go first: gpt-5.2 retires on 8 June 2027 and gpt-5.1 on 13 July, so neither is a landing spot for GPT-5.4 traffic. For the February 2027 GPT-5 wave, see our Azure GPT-5 retirement checklist.

The 12-month cutoff most admins miss

Foundry's lifecycle has five stages: Preview, GA, an optional Legacy stage, Deprecated and Retired. A Deprecated model keeps serving existing deployments, and existing customers can still create and manage deployments, but new customers can neither deploy nor access it.

Microsoft decides who counts as an existing customer at the subscription level: whether that Azure subscription has ever deployed the specific model version. A new subscription under the same tenant does not inherit access. Three common situations hit this rule:

• New landing zones. A platform team that splits workloads into new subscriptions in mid-2027 cannot deploy gpt-5.4 there once it is Deprecated.

• Disaster recovery. A DR subscription that has never deployed gpt-5.5 cannot stand it up during an incident after about April 2027.

• Per-customer subscriptions. SaaS vendors that give each enterprise customer its own subscription cannot onboard a new customer onto gpt-5.4 after the cutoff.

Before March 2027, deploy the version once in every subscription that may need it until retirement, or decide now that those subscriptions start on the successor. Check status through the Models API, a subscription-scoped call per region:

            GET https://management.azure.com/subscriptions/{sub}/providers/Microsoft.CognitiveServices/locations/{location}/models?api-version=2024-10-01
          

Read lifecycleStatus, deprecation.inference and the per-SKU deprecationDate. The API uses different words from the documentation: Deprecating means what the docs call Deprecated (still serving, closed to new customers), and Deprecated means retired. Microsoft treats any past deprecation.inference date as retired, whatever the status field says.

Dates and behaviour by deployment type

Unlike gpt-4o version 2024-05-13, whose Regional Standard deployments retired early, on 31 March 2026, the GPT-5.4 and GPT-5.5 rows carry no per-offer footnote. Plan on one date for every deployment type, and confirm it against the per-SKU deprecationDate.

What differs is where each model is offered and what happens on retirement day. This table uses the Americas tab of Microsoft's region availability page on 7 November 2026; other tabs list different regions.

Three gaps stand out. gpt-5.4-pro and gpt-5.4-nano appear in no provisioned table, so most teams have no reserved capacity to unwind for them. No GPT-5.6 or GPT-6 model appears in either Batch table yet, which leaves a gpt-5.4 Batch workload without a newer Batch target. And Regional Provisioned lists GPT-5.6 only in West US (Sol, Terra) and East US 2 (Luna), with no GPT-6 model, so a regional PTU deployment of gpt-5.4 in Central US or gpt-5.5 in East US has no in-region provisioned successor yet.

Auto-upgrade rules, and when to switch them off

For Global Standard, Data Zone Standard and Standard deployments, Microsoft upgrades retiring versions automatically, region by region, even where the new version is not yet separately offered in that region or SKU. Priority Processing follows the same path. Provisioned deployments are never auto-upgraded, and Batch uses a side-by-side move.

The deployment property versionUpgradeOption controls Standard behaviour:

You can read and change the value through REST, Azure PowerShell or the Foundry portal; the Azure CLI can read it but cannot update it yet. A minimal PowerShell change, adapted from Microsoft's own steps:

            $d = Get-AzCognitiveServicesAccountDeployment -ResourceGroupName rg-ai -AccountName aoai-prod -Name gpt54-agent
$d.Properties.VersionUpgradeOption = "NoAutoUpgrade"
New-AzCognitiveServicesAccountDeployment -ResourceGroupName rg-ai -AccountName aoai-prod -Name gpt54-agent -Properties $d.Properties -Sku $d.Sku
          

A blank Replacement column is a risk. The schedule calls its Replacement column the automatic-upgrade target only where upgrades are supported, while the model-update page says a pinned version moves to the default version at retirement. Neither says plainly what a Standard deployment does if the column is still empty on the day. Plan as if it simply stops.

When to switch auto-upgrade off. Choose NoAutoUpgrade for regulated workflows, tool-calling agents and strict output schemas, because an automatic swap changes behaviour without a test run. The cost: a forgotten deployment returns 410 Gone on retirement day.

Provisioned throughput. The replacement should reach provisioned regions about 30 days before retirement: about 3 August 2027 for gpt-5.4 and 26 September for gpt-5.5. An in-place migration keeps the deployment name and size while Azure moves traffic over 20 to 30 minutes; it works only between provisioned deployments of the same type, and a model family change needs the REST API or CLI. A multi-deployment migration creates a new deployment, possibly of another type, shifts traffic, waits until the old one shows no requests for 5 to 10 minutes, then deletes it. PTU capacity is fungible across provisioned deployments, but confirm quota and regional capacity first.

What replaces GPT-5.4 and GPT-5.5

Microsoft declares the official replacement 90 to 120 days before retirement so the pick reflects the best model available then, and tells customers not to wait for it. These are the candidates already on Azure, at Global Standard short-context rates from the Azure Retail Prices API (East US 2 meters) on 7 November 2026.

Watch the Sol price. OpenAI says its $4 and $20 rate for GPT-5.6 Sol is promotional until 30 November 2026. Azure's meters show the same figures with no end date, but if Azure returns to $5 and $30, gpt-5.6-sol costs exactly what gpt-5.5 costs today.

The GPT-6 family (gpt-6-astra from 2026-09-03, gpt-6-sol and gpt-6-luna from 2026-09-22, gpt-6.1-sol from 2026-09-29, gpt-6.1-terra from 2026-10-26) is on Microsoft's models page but not yet on the retirement schedule. Under the 18-month rule those versions would run to March or April 2028. Some quota tiers must request quota for gpt-5.5, GPT-5.6 and GPT-6; Tier 5 and Tier 6 subscriptions have it by default.

Code changes to test before any switch

• Tools on Chat Completions. Microsoft says the gpt-5.6 models support Chat Completions and function tools, but not both in one request unless reasoning_effort is none, and points to the Responses API for tool calling; it lists gpt-6.1-sol tool calling on the Responses API only. gpt-5.4, gpt-5.4-mini, gpt-5.4-nano and gpt-5.5 accept tools on Chat Completions today, so an auto-upgrade can break a working agent.

• Context windows. gpt-5.4, gpt-5.4-pro, gpt-5.5 and the gpt-5.6 models have a 1,050,000-token window with 922,000 input; gpt-5.4-mini and gpt-5.4-nano have 400,000 with 272,000 input.

• Long-context pricing. Azure meters requests above 272,000 input tokens separately. Long-context gpt-5.4 costs $5 input and $22.50 output; gpt-5.6-sol costs $8 and $30.

• Cache writes. GPT-5.6 and GPT-6 meters add a cache-write charge that GPT-5.4 and GPT-5.5 meters lack, so prompt-caching savings need re-measuring.

Worked cost comparison

Illustrative arithmetic at the Global Standard rates above, with no caching and the same token counts before and after. eCorpIT did not benchmark these models; reasoning tokens bill as output and vary by model and effort.

GPT-6 Sol would take workload A to $6,000 + $4,000 = $10,000 (-26%), though it has no retirement date yet. The real risk is quality and token volume: a cheaper model that needs higher effort to match today's accuracy can erase the saving, so compare cost per completed task on your own traces.

Test and rollout plan

Dates assume Microsoft keeps its windows and use gpt-5.4's dates for the family; gpt-5.4-pro runs 5 days later, mini and nano 19 days later. The 90-day and 30-day markers are when the replacement should reach Global Standard and provisioned regions; the 60-day marker is the latest date for Microsoft's retirement email.

The steps behind that table:

1. Inventory and map. List model, version, deployment type, region, subscription and versionUpgradeOption for all five models, and tie each deployment to its applications, agents and gateway routes.

1. Set alerts. In Azure Service Health, filter health advisories by Azure OpenAI Service, the name Microsoft uses for Foundry there.

1. Move tool calls to the Responses API while still on gpt-5.4 or gpt-5.5, so the API change and the model change ship separately.

1. Build evaluation sets. Capture 200 to 500 recent traces per workflow, label 50 to 200 golden cases, and record baseline quality, p95 latency and cost per task.

1. Score two candidates per workload at two or three effort levels against pass marks set in advance; our guide to catching silent agent failures in CI/CD covers the gate.

1. Canary through a gateway. An AI gateway such as LiteLLM, Portkey, Kong or TrueFoundry makes each step a configuration change, keeps the old deployment as rollback, and lets you alert on any 410 Gone after each date.

India, EU and US notes

Processing location. Global deployments may process prompts in any Azure region where the model is deployed, Data Zone deployments anywhere inside the US, EU or APAC zone, and regional deployments in the deployment region; data at rest stays in the designated Azure geography. If a DPDP, GDPR or sector assessment assumed a processing location, re-check it whenever a migration changes deployment type.

India. All five retiring models and every GPT-5.6 and GPT-6 model are listed under Global Standard in South India. Under APAC Data Zone Standard, gpt-5.4 and gpt-5.4-mini appear in all five APAC regions but gpt-5.5 and gpt-5.4-nano do not, and of the newer models only gpt-5.6-sol, gpt-6.1-sol and gpt-6.1-terra appear. South India hosts gpt-5.4-mini under Regional Provisioned, with no GPT-5.6 or GPT-6 model listed there. Microsoft's Voice Live API lists gpt-5.4 as a Pro-tier model; our guide to the June 2027 gpt-realtime retirement on Azure covers the voice side.

EU and US Data Zones. Data Zone gpt-5.4 output costs $16.50 per million tokens against $15 for Global, a 10% premium, in both the East US 2 and Sweden Central meters, and GPT-5.6 carries the same 10%. GPT-6 differs: its Data Zone premium is 10% in East US 2 but 20% in Sweden Central, where gpt-6-sol output costs $12 against $10 Global. Budget for the uplift on the successor.

Government clouds. Azure Government offers no Global Standard and usually supports one version of a model at a time, with a 30-day overlap, leaving little room for side-by-side tests.

How eCorpIT can help

eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our LLM migration and cost optimisation service inventories Azure OpenAI deployments, scores candidate models on your own traces and plans provisioned and Batch moves before each retirement date. Talk to us at /contact-us/.

Last updated: 7 November 2026.

References

1. Model retirement schedule, Microsoft Foundry, last updated 23 September 2026 — Microsoft Learn

1. Foundry Models lifecycle and support policy — Microsoft Learn

1. Azure Retail Prices API, GPT-5.4 Global meters in East US 2 — Microsoft

1. Azure Retail Prices API, GPT-5.5 meters in East US 2 — Microsoft

1. Azure Retail Prices API, GPT-5.6 Global meters in East US 2 — Microsoft

1. Pricing — OpenAI API

1. Deprecations — OpenAI API

1. Region availability for Foundry Models sold by Azure — Microsoft Learn

1. Azure OpenAI in Microsoft Foundry Models: working with models — Microsoft Learn

1. Azure Retail Prices API, Azure OpenAI GPT6 meters in East US 2 — Microsoft

1. Foundry Models sold by Azure — Microsoft Learn

1. Azure OpenAI Service pricing — Microsoft Azure

1. Voice Live API overview — Microsoft Learn

1. Azure Retail Prices API, GPT-5.4 Data Zone meters in East US 2 — Microsoft

1. Azure Retail Prices API, GPT-5.4, GPT-5.6 and GPT-6 output meters in Sweden Central — Microsoft