Azure OpenAI Codex retirement: 6 models from 17 March to 24 August 2027

Summary. Microsoft's model retirement schedule lists six Codex models leaving Azure OpenAI in 2027: gpt-5-codex (version 2025-09-15) on 17 March, gpt-5.1-codex and gpt-5.1-codex-mini (2025-11-13) on 15 May, gpt-5.1-codex-max (2025-12-04) on 18 May, gpt-5.2-codex (2026-01-14) on 13 July and gpt-5.3-codex (2026-02-24) on 24 August. On 7 November 2026 all six are GA with no replacement named; Microsoft picks one roughly 90 to 120 days before retirement, so the gpt-5-codex pick should land between 17 November and 17 December 2026. Standard deployment types then upgrade to it unless set to NoAutoUpgrade; Provisioned deployments never do, and a retired model returns 410 Gone. Codex CLI and CI jobs call a deployment name, so an auto-upgrade swaps the model underneath without any config change. Azure Global Standard prices run from $0.25 input and $2 output per million tokens (gpt-5.1-codex-mini) to $1.75 and $14 (gpt-5.2-codex and gpt-5.3-codex), against $2 and $10 for GPT-6.1 Sol. Choose your target and upgrade policy before mid-December.

About this article. This is for platform teams running Codex CLI, CI jobs or internal coding agents on Azure OpenAI in Microsoft Foundry. We checked every date, price and rule on 7 November 2026 against Microsoft Learn's retirement schedule, lifecycle policy, model and region pages and Codex guide, the Azure Retail Prices API, and OpenAI's deprecations, model and Codex documentation. 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 advice, and no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.

The six Azure retirement dates

Microsoft's retirement schedule, last updated on 23 September 2026 and unchanged when we read it on 7 November, gives the rows below. Prices are Azure Global Standard rates per million tokens for East US 2 from the Azure Retail Prices API.

The older codex-mini (2025-05-16), a fine-tuned o4-mini, is already Deprecated and retires on 15 November 2026 with no replacement listed, eight days from now.

Three details on the schedule are worth knowing:

• Most dates follow the 18-month rule; one does not. Microsoft's lifecycle policy sets GA retirement 18 months after launch. Counted from the version date, gpt-5.3-codex lands exactly on 18 months and four others within two days. gpt-5.1-codex-max retires 17 days before its 18-month mark of 4 June.

• One version string differs between pages. The schedule and region tables list gpt-5-codex as 2025-09-15; the capabilities table on the Foundry Models sold by Azure page lists 2025-09-11. Check what your deployment reports.

• Lifecycle labels lag the policy. At 12 months a GA model becomes Deprecated: subscriptions that never deployed that version cannot deploy it, including new subscriptions in the same tenant. gpt-5-codex passed 12 months on 15 September 2026 yet still shows GA; the gpt-5.1-codex models reach 12 months on 13 November and 4 December. Test a fresh deployment in a new subscription or disaster-recovery region now.

Microsoft says retirement dates cannot be extended.

Azure versus the OpenAI API

Azure now outlives the OpenAI API for all six. OpenAI shut down gpt-5-codex, gpt-5.1-codex, gpt-5.1-codex-max, gpt-5.1-codex-mini and gpt-5.2-codex on 23 July 2026, and removes gpt-5.3-codex on 1 April 2027 with gpt-6-sol as its recommended replacement. Our guide to the OpenAI API shutdown of GPT-5.3-Codex, GPT-5.1 and GPT-5.4 nano covers that side.

Where an agent pairs OpenAI with Azure, the Azure half lives longer; test it as the fallback.

When Microsoft should name replacements

The lifecycle policy says the official replacement is chosen roughly 90 to 120 days before retirement, becomes testable in Global Standard about 90 days before, and reaches the provisioned regions where the old model is retiring about 30 days before. GA retirement notices go by email to subscription owners with active deployments at least 60 days ahead.

A provisioned team that waits for Microsoft's pick gets about 30 days to test it. Microsoft advises against waiting: the listed replacement is its recommendation and auto-upgrade target, not a limit on which model you move to by hand. Our comparison of model retirement policies sets this cycle against other vendors.

Auto-upgrade behaviour by deployment type

These rules come from the lifecycle policy page, last updated on 24 July 2026.

The versionUpgradeOption property takes OnceNewDefaultVersionAvailable, OnceCurrentVersionExpired or NoAutoUpgrade, set through REST, Azure PowerShell or the Foundry portal. Microsoft also warns that an upgrade can happen even where the new version is not yet separately offered in that region or SKU.

Where the six models can be deployed

Microsoft's region availability page, updated on 19 October 2026, shows a footprint that narrows quickly outside Global Standard:

• Global Standard. All six are listed across the Americas, Asia Pacific (including South India) and Middle East and Africa regions. In Europe, gpt-5-codex and the gpt-5.1-codex family are missing from North Europe.

• Data Zone Standard. Only gpt-5.3-codex is listed, in the US and Asia Pacific zones. None of the six appears in the EU zone.

• Provisioned. Global Provisioned lists gpt-5.1-codex, gpt-5.2-codex and gpt-5.3-codex only. Data Zone Provisioned lists only gpt-5.1-codex (US and EU), and Regional Provisioned only gpt-5.3-codex in UK South.

• Regional Standard and Batch. None of the six is listed.

Global deployments may process prompts in any Azure region where the model is deployed; Data Zone deployments stay within the US, the EU or Asia Pacific, which is not an India-only zone. A team with an EU or India processing rule that runs Codex models on Global Standard should check whether that ever met its rules. The GPT-6 models appear in the US and EU Data Zone Standard tables, and gpt-6.1-sol and gpt-6.1-terra in the Asia Pacific one, so a migration can close a residency gap. One catch: gpt-6.1-terra is not listed for East US 2 in Global Standard or the US data zone, and has no provisioned listing.

What the target models cost on Azure

Azure Global Standard, East US 2, short-context rates from the Azure Retail Prices API on 7 November 2026, per million tokens:

Long-context requests bill higher, for example $4 input and $15 output on gpt-6.1-sol. Microsoft says the number of input tokens in a request decides whether GPT-6 Standard deployments bill it as short or long context, and that some quota tiers must request GPT-6 quota, while Tier 5 and Tier 6 subscriptions have it by default. On the standard 18-month lifecycle the GPT-6 models would retire around March and April 2028, but the schedule does not list them yet.

Worked monthly cost

Illustrative arithmetic for one coding-agent workload: 2,000 million input tokens a month, 80% from cache (1,600 million), 400 million uncached and 60 million output. For the newer models the range covers uncached input billed at the base rate or entirely as cache writes. Token counts are held constant, which they will not be in practice, and eCorpIT did not benchmark these models. Size your own volumes with our LLM token counter.

GPT-6.1 Sol line: 1,600 × $0.10 + 400 × $2 (or $2.50) + 60 × $10. For a lighter gpt-5.1-codex-mini workload of 500 million input tokens (400 million cached) and 20 million output, today's $75 a month (400 × $0.025 + 100 × $0.25 + 20 × $2) becomes $24 to $26.50 on gpt-6-luna and $220 to $245 on gpt-6.1-terra. In the larger example, GPT-6.1 Sol costs 20% to 35% more than the older Codex models and 3% to 14% less than gpt-5.2-codex or gpt-5.3-codex.

What coding-agent teams must change

Codex CLI and the VS Code extension

Microsoft's Codex on Azure guide configures Codex against the v1 Responses API, and the model value is your Azure deployment name, not OpenAI's model ID:

            model = "gpt-5-codex"  # your Azure deployment name
model_provider = "azure"
model_reasoning_effort = "medium"

[model_providers.azure]
name = "Azure OpenAI"
base_url = "https://YOUR_RESOURCE_NAME.openai.azure.com/openai/v1"
env_key = "AZURE_OPENAI_API_KEY"
wire_api = "responses"
          

If the deployment named gpt-5-codex is auto-upgraded, every developer silently gets a different model; if it is set to NoAutoUpgrade, every session fails on 17 March. Handle either outcome on purpose:

1. Create a new deployment for the target model with a descriptive name, such as coding-gpt-6-1-sol, and keep the old one until cut-over.

1. Distribute the change centrally. OpenAI's managed configuration sets Codex requirements and defaults, and teams using an Azure key rather than a ChatGPT sign-in can serve them from their own HTTPS endpoint.

1. Check reasoning effort values. On Azure, gpt-5.1-codex-max accepts xhigh but not none. OpenAI lists low to max for GPT-6.1 Sol, with no none, and low to xhigh for GPT-5.3-Codex. Unsupported values fail.

1. Keep keys in environment variables. Microsoft says env_key must name a variable rather than hold the key, and that Codex does not yet support Entra ID.

GitHub Actions and PR bots

Microsoft's guide runs Codex in GitHub Actions with codex -p azure exec --full-auto and the key in a repository secret. Its sample config defines no profile, and from Codex 0.134.0 a profile lives in its own file, such as ~/.codex/azure.config.toml, so check that the named profile exists on the runner. OpenAI's Codex GitHub Action (openai/codex-action@v1) exposes model, effort and codex-version inputs; a blank codex-version installs the latest release. For each workflow:

• Set the deployment name explicitly.

• Pin codex-version so a CLI upgrade and a model change are separate events. OpenAI says GPT-6.1 Terra needs Codex CLI 0.159.0 or later, and Microsoft lists 0.155.1 as validated with gpt-6-sol and gpt-6-luna on Azure.

• Run the new deployment on a copy of the review workflow for two weeks, compare results, then switch.

OpenAI's configuration reference lists responses as the only supported wire_api value, so any proxy in front of Azure must pass the Responses API through.

Internal tools and agent frameworks

Review bots, migration scripts and agent frameworks call deployments through the Azure OpenAI SDKs, so search for deployment names as well as model IDs. None of the six Codex models lists Chat Completions on Azure, so existing code already uses the Responses API. Stay there: gpt-6.1-sol supports tool calling only through Responses. Context grows from 400,000 tokens (272,000 input) to 1,050,000 (922,000 input), and long-context requests cost more, so cap agent context deliberately.

Monitoring

Use the Models API per region to read lifecycleStatus, deprecation.inference and the per-SKU deprecationDate:

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

In the API, Deprecating means the docs' Deprecated, and Deprecated means retired. In Azure Service Health, filter advisories by Azure OpenAI Service, as there is no Foundry entry.

Checklist for coding-agent teams

1. Inventory deployments. List every deployment of the six models and codex-mini across subscriptions and regions, with version, deployment type and versionUpgradeOption.

1. Map callers. Tie each deployment name to developer configs, CI workflows, bots and internal services.

1. Set the upgrade policy. For coding agents, set NoAutoUpgrade and migrate on your own schedule, with an alert so nothing reaches 410 Gone unnoticed.

1. Choose targets now. Shortlist two, typically GPT-6.1 Sol and GPT-6.1 Terra, and confirm region, deployment type and quota for each.

1. Build a golden set. Take 50 to 200 real tasks with tests per repository group, record pass rate, tokens, tool-call errors and wall time on the retiring model, and set pass gates before testing candidates.

1. Plan PTU moves. Confirm quota and choose in-place or side-by-side migration.

1. Cut over in stages. Move 10%, 50% and then 100% of developers or workflows, with the old deployment as rollback until its date.

1. Retire configs. Remove old deployment names from managed configuration.

Five retirement dates in just over five months argue for one migration, not five: moving every Codex deployment to one current target by February clears the list. gpt-5 and gpt-5-nano go earlier, on 9 February 2027; our Azure OpenAI GPT-5 retirement checklist covers them.

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 AI coding agent rollout and governance service inventories Azure OpenAI deployments and the tools that call them, runs golden-set evaluations of candidate models on your repositories and sets up managed Codex configuration. Talk to us at /contact-us/.

Last updated: 7 November 2026.

References

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

1. Foundry Models lifecycle and support policy, updated 24 July 2026 — Microsoft Learn

1. Codex with Azure OpenAI in Microsoft Foundry Models — Microsoft Learn

1. Azure Retail Prices API, Foundry Models meters, read 7 November 2026 — Microsoft

1. Foundry Models sold by Azure, updated 29 October 2026 — Microsoft Learn

1. Deprecations — OpenAI API

1. Region availability for Foundry Models sold by Azure, updated 19 October 2026 — Microsoft Learn

1. Managed configuration — OpenAI

1. GPT-6.1 Sol model — OpenAI API

1. GPT-5.3-Codex model — OpenAI API

1. Advanced configuration, Codex — OpenAI

1. Codex GitHub Action — OpenAI

1. Models, Codex — OpenAI

1. Configuration reference, Codex — OpenAI