Azure OpenAI GPT-5 retirement on 9 February 2027: upgrade checklist

Summary. Microsoft's retirement schedule lists gpt-5 and gpt-5-nano, version 2025-08-07, retiring on Azure on 9 February 2027, after which requests return 410 Gone. gpt-5-codex follows on 17 March 2027, gpt-5-pro on 7 April 2027 and gpt-5-mini on 13 July 2027. On 1 November 2026 the Replacement column is blank for every GPT-5 model, although Microsoft names replacements 90 to 120 days before retirement, a window running from 12 October to 11 November for the February date. Provisioned deployments are never auto-upgraded. The OpenAI API is earlier: it shuts down gpt-5, gpt-5-mini, gpt-5-nano and gpt-5-pro on 11 December 2026 and recommends gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna, which Azure keeps until 11 January 2028. Azure lists GPT-5 at $1.25 input and $10 output per million tokens, GPT-6 Sol at $2 and $10, and GPT-5.6 Sol at $4 and $20. gpt-5-nano at $0.05 and $0.40 has no like-for-like successor: GPT-6 Luna costs $0.10 and $0.50. An upgrade to gpt-5.6 can also break Chat Completions tool calls. Pick your target now, and expect Microsoft's 60-day notice by 11 December.

About this article. This checklist is for IT admins and engineering leads running GPT-5 models on Azure OpenAI in Microsoft Foundry. Dates, prices and rules were checked on 1 November 2026 against Microsoft Learn's Foundry pages, the Azure OpenAI pricing page, the Azure Retail Prices API, Microsoft's GPT-6 launch post and OpenAI's deprecations and model pages. We read the published pages and did not query the Models API from a live subscription. Vendors change schedules and prices often, so re-check before you commit; this is not procurement or legal advice, and no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.

What retires on Azure, and when

Microsoft's model retirement schedule, last updated on 23 September 2026, and OpenAI's deprecations page, whose GPT-5 notice is dated 11 June 2026, give the columns below.

Azure outlives OpenAI by two months. Code that calls both loses the OpenAI side on 11 December 2026, and an Azure fallback disappears on 9 February.

There is no fine-tuning tail. The schedule's fine-tuned table covers gpt-4o, gpt-4.1 and o4-mini, not GPT-5, and OpenAI's GPT-5 pages list fine-tuning as not supported.

gpt-5.1 is a poor landing spot. It retires on Azure on 13 July 2027 and on the OpenAI API on 1 April 2027. It is also Azure's auto-upgrade target for gpt-4o, which our GPT-4o and GPT-4.1 shutdown guide covers. Moving GPT-5 traffic there buys five months at most.

The gpt-5-mini outlier and other page discrepancies

Microsoft publishes the rules on one page and the dates on another, and they disagree. The lifecycle policy was last updated on 24 July 2026 and the schedule on 23 September 2026, though that stamp lags: the schedule also records a deprecation dated 9 October 2026.

The gpt-5-mini date matters most. It shares the 2025-08-07 version with gpt-5 and gpt-5-nano, yet neither page explains the later date, so treat it as something that could move. Microsoft's FAQ tells customers to plan with the schedule and the Models API. Views can differ: a March 2026 thread on Microsoft Q&A describes the portal showing 5 February 2027 on a gpt-5 deployment. Read the date on your own deployments.

The Models API is the programmatic check, 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. In the API, Deprecating means what the docs call Deprecated, and Deprecated means retired. Microsoft treats any past deprecation.inference date as retired, whatever the status field says.

Auto-upgrade behaviour by deployment type

Microsoft upgrades Standard deployment types automatically at retirement, rolling region by region, and the schedule's Replacement column names the target where that is supported. With that column blank for GPT-5, read this table as what would happen once a replacement is named. Recent picks differ: gpt-5.1 for gpt-4o, gpt-5.6-sol or gpt-5.6-terra for the o-series.

The versionUpgradeOption property takes three values, set through REST, Azure PowerShell or the Foundry portal. OnceCurrentVersionExpired upgrades at retirement, as does a deployment with no value set. OnceNewDefaultVersionAvailable upgrades within two weeks of a new default version. NoAutoUpgrade never upgrades.

For a regulated workload, NoAutoUpgrade plus a planned migration is safer, because an automatic swap changes behaviour without a test run; Microsoft's migration guide calls auto-upgrade a safety net, not a plan. The trade-off is that a forgotten deployment returns 410 Gone on 9 February.

Provisioned. Microsoft says the replacement reaches Global Standard about 90 days before retirement and provisioned regions about 30 days before, so waiting on its pick means a test window from about 10 January. PTU capacity is fungible across provisioned managed deployments, but the target must support provisioned in your region. The launch post lists Provisioned Throughput for GPT-6 Astra and Sol; the current region table also shows gpt-6-luna under Global Provisioned and Data Zone Provisioned, though the pricing page has no provisioned row for Luna. gpt-5-nano appears under no provisioned type.

Batch. Microsoft's Global Batch and Data Zone Batch tables list gpt-5 but not gpt-5-mini or gpt-5-nano. The only later GPT-5 family models they list are gpt-5.1, gpt-5.4 and gpt-5.4-mini, and no GPT-5.6 or GPT-6 model appears. A gpt-5 Batch workload therefore lands on gpt-5.4 or gpt-5.4-mini, which retire in September 2027, or moves to synchronous calls.

Replacement mapping: what to move to

OpenAI's mapping is explicit. Microsoft's launch post only advises customers on legacy models to evaluate GPT-5.6 Sol or newer, and its lifecycle page says not to wait for a named replacement.

Prices are per million tokens at Azure Global Standard, short context, from the Azure pricing page and Retail Prices API; the GPT-5 family matches OpenAI's list prices. GPT-6 models are not on the retirement schedule yet.

OpenAI says GPT-5.6 Sol's $4 and $20 are promotional, available at least through 21 November 2026, and represent cuts of 20% and 33%, which implies $5 and $30 if the offer ends. Azure's page shows $4 and $20 with no end date, so re-check after 21 November; our GPT-5.6 Sol pricing analysis has the history.

OpenAI maps the GPT-5.6 tiers to the old ones (Sol to the unsuffixed tier, Terra to mini, Luna to nano), but per-token prices did not follow. Against gpt-5-mini and gpt-5-nano, context grows from 400,000 tokens to 1,050,000, the knowledge cutoff moves from 31 May 2024 to 16 February 2026, and prompts over 272,000 input tokens cost 2x input and 1.5x output for the whole request. gpt-5.6 also bills cache writes at 1.25x input, a line the GPT-5 pages do not carry.

Code changes to test

• Tools on Chat Completions. Microsoft warns that gpt-5.6 models reject a Chat Completions request that sends tools unless reasoning_effort is none, even if your code never sets it, because they default to medium, so apps can fail after a deployment upgrade. OpenAI's GPT-6 Sol and Luna pages state the same limit, and GPT-6.1 Sol and Terra require the Responses API for tool calling. Move tool-using agents to the Responses API.

• Effort values. minimal works only on the original GPT-5 models, not on gpt-5.1 or later. Set effort explicitly on every call.

• Long conversations. gpt-5.6 and GPT-6 models render earlier turns' reasoning items by default, which raises billed tokens with no code change. On gpt-5.6, setting reasoning.context to current_turn keeps the earlier behaviour.

Our GPT-4o migration guide covers prompt and structured-output changes. If you are weighing Claude on Foundry, claude-opus-4-6 and claude-sonnet-4-6 retire there on 2 and 10 February 2027; see our Claude retirement guide.

Worked cost comparison

This is illustrative arithmetic from the Azure Global Standard prices above, with no caching and the same token counts before and after. eCorpIT did not benchmark these models, and reasoning-token volumes differ by model and effort.

OpenAI's mapping raises costs steeply. GPT-6 Sol adds 20%, while GPT-6 Luna saves 69% against gpt-5-mini but costs 61% more than gpt-5-nano. GPT-6.1 Terra sits between them in Scenario C at $6,000 + $5,000 = $11,000 (+214%). If gpt-5.6-sol moved to the implied $5 and $30, Scenario B would reach $60,000 + $90,000 = $150,000 (+233%).

A model that finishes the task in fewer reasoning tokens can close the gap, which is why Microsoft argues for cost per task rather than per token. Measure that on a golden set, and use our LLM token counter to rerun the numbers.

IT admin checklist: 100 days to 9 February

From 1 November 2026 there are 100 days to 9 February 2027.

1. Inventory every GPT-5 deployment. List deployments across subscriptions and regions with model, version, deployment type and versionUpgradeOption, including gpt-5-codex and gpt-5-pro.

1. Find the callers. Map each deployment name to the applications, agents, Logic Apps and gateway routes that use it.

1. Set alerts. In Azure Service Health, filter health advisories by Azure OpenAI Service, the name Microsoft uses for Foundry, and route them to email or webhook.

1. Decide the upgrade policy. Set NoAutoUpgrade on regulated workloads and plan a manual cut-over. Elsewhere, decide whether to accept Microsoft's pick, and re-read the schedule after 11 November.

1. Choose targets per workload. Shortlist two candidates and confirm deployment-type and regional availability, quota and content-filter configuration.

1. Build the eval set. 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. Run candidates at two or three effort levels. Score task success, schema validity, tool-call accuracy and refusals against gates set in advance.

1. Plan provisioned and Batch moves. Confirm PTU quota, choose in-place or side-by-side migration outside peak hours, and move Batch jobs to a model the Batch tables list.

1. Canary and cut over. Route 5%, 25%, then 100% of traffic through your gateway, keeping a second current model as rollback.

1. Watch for 410s. After 9 February, alert on any 410 Gone from a GPT-5 deployment name.

An AI gateway such as LiteLLM, Portkey, Kong or TrueFoundry makes each cut-over a configuration change, and our guide to catching silent agent failures in CI/CD covers the eval gate.

India, EU and US notes

India. Global deployments can process prompts in any Azure region where the model is deployed, and Data Zone deployments anywhere in the zone. South India appears in the Asia Pacific Data Zone Standard table for gpt-5.6-sol, gpt-6.1-sol and gpt-6.1-terra, but not for gpt-5 or gpt-6-luna. Only regional types keep processing in your chosen geography. South India lists gpt-5 and gpt-5-mini under Regional Provisioned, with gpt-5.1 and gpt-5.4-mini the later-dated options; no GPT-5.6 or GPT-6 model appears there, and no GPT-5 or later model is listed under regional Standard in India. Recheck any DPDP assessment that assumed in-country processing, keep prompt logging minimal and record the processing location.

EU. GPT-5 Data Zone prices carry a 10% premium, while GPT-6 EU Data Zone prices carry 20%: Luna $0.12 and $0.60, Sol $2.40 and $12. On the Data Zone, Scenario A goes from $1,705 to $3,000 (+76%) and Scenario B from $49,500 to $64,800 (+31%).

US. The US Data Zone carries a 10% premium for GPT-6 too, so Scenario B on GPT-6 Sol at $2.20 and $11 costs $59,400 (+20% on $49,500). Government clouds usually support one version of a model at a time, with a 30-day overlap, and offer no Global Standard.

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 your GPT-5 deployments, scores candidate models on your production traces and plans the provisioned and Batch moves. Talk to us at /contact-us/.

Last updated: 1 November 2026.

References

1. Model retirement schedule, Microsoft Foundry — Microsoft Learn

1. Foundry Models lifecycle and support policy — Microsoft Learn

1. Deprecations — OpenAI API

1. Azure OpenAI pricing — Microsoft Azure

1. Azure OpenAI reasoning models — Microsoft Learn

1. GPT-5 model — OpenAI API

1. Clarification on OAI model expiration date, March 2026 — Microsoft Q&A

1. Working with Azure OpenAI models — Microsoft Learn

1. Model migration process — Microsoft Learn

1. GPT-6 Astra, Sol, and Luna for production AI agents in Microsoft Foundry, 22 September 2026 — Microsoft Azure Blog

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

1. Azure Retail Prices API, Foundry Models meters — Microsoft

1. GPT-5 mini model — OpenAI API

1. GPT-5 nano model — OpenAI API

1. GPT-5.6 Sol model — OpenAI API

1. GPT-5.6 Terra model — OpenAI API

1. GPT-5.6 Luna model — OpenAI API

1. GPT-6 Sol model — OpenAI API

1. GPT-6 Luna model — OpenAI API

1. GPT-6.1 Sol model — OpenAI API

1. GPT-6.1 Terra model — OpenAI API

1. Deployment types for Microsoft Foundry Models — Microsoft Learn