Azure OpenAI and Foundry Retirements in November 2026: o1, o1-pro, codex-mini, Claude Haiku 4.5

Last updated: October 18, 2026

Microsoft Foundry (the platform that now hosts Azure OpenAI) has three retirement dates in one week of November. On Sunday, November 15, 2026, codex-mini and both versions of claude-haiku-4-5 retire. On Tuesday, November 17, two preview versions of gpt-chat-latest go. On Thursday, November 19, o1, o1-pro, o3, o3-pro, o3-deep-research, o3-mini and o4-mini all retire, with gpt-5.6-sol or gpt-5.6-terra named as the replacement. The first of those dates is 28 days away. This post covers what each date means for Standard and Provisioned deployments, the replacements Microsoft names (and ours where it names none), prices, and the deployment change. Everything below was checked against Microsoft's own docs on October 18, 2026.

The dates#

All rows come from Microsoft's model retirement schedule, last updated September 23, 2026.

Two more Claude dates follow in the same month on the same page: claude-opus-4-5 on November 24 (no replacement listed) and claude-sonnet-4-5 on November 30 (replacement claude-sonnet-5-5). The Sonnet date is covered in our shutdown calendar because it lands on the Claude API the same day.

Deployment types. The schedule gives one date per model version. It does not list separate dates for Global Standard, Data Zone, Standard or Provisioned deployments. What differs by deployment type is what happens on that date, covered in the next section. Microsoft's lifecycle policy also notes that auto-upgrades "are scheduled on a rolling, region-by-region basis", so the switch-over may not land at the same hour everywhere.

Fine-tuned o4-mini is the one exception. Fine-tuned models retire in two phases, training and deployment, and fine-tuned o4-mini deployments outlive the base model. The schedule page says deployment retirement is November 19, 2027, with training no earlier than May 19, 2027. The lifecycle policy page has an older table that says October 16, 2027 and April 16, 2027. Plan against the earlier pair until Microsoft reconciles them.

Our Azure OpenAI deprecations page tracks every Foundry row, and the full deprecations table puts them next to OpenAI, Anthropic, Bedrock and Vertex.

Deprecated vs retired, and what happens on the date#

Microsoft uses five lifecycle stages. Two matter here:

• Deprecated: the model still serves traffic. Existing customers can "continue to create and manage deployments", but new customers cannot. "Existing customer" is decided per Azure subscription, by whether that subscription "has ever deployed the specific model version". A new subscription in the same tenant does not inherit access. codex-mini and all seven o-series models are in this stage now.
• Retired: "Removed from service. All inference requests return 410 Gone."

One trap if you check status in code: the Models API uses different words. A model the docs call Deprecated shows lifecycleStatus: "Deprecating", and the API value "Deprecated" means it is already retired. Microsoft's policy page spells this out and gives the endpoint:

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

Retirement dates cannot be moved. The policy FAQ answers "Can I get an exception to extend a model's retirement date?" with "No. Retirement dates aren't extendable."

Standard deployments: auto-upgrade, if there is a target#

For Global Standard, Data Zone Standard and Standard deployments, Microsoft runs automatic upgrades at retirement. Each deployment has a versionUpgradeOption, documented on the working with models page:

The schedule's Replacement column is "the target for automatic upgrades at retirement" where supported. So for the November 19 rows, a Standard o1 or o3 deployment that never touched this setting is expected to become a gpt-5.6-sol deployment on that date, and o3-mini or o4-mini becomes gpt-5.6-terra. Microsoft's worked example is gpt-4o 2024-05-13 retiring on December 9: "the service automatically upgrades it to gpt-5.6-sol on the Standard SKU in each region where that version is currently available." Quota carries over automatically for Standard auto-upgrades, per the policy FAQ. Priority Processing upgrades the same way as Standard.

An auto-upgrade can still break your app. Microsoft's reasoning models page warns that the gpt-5.6 models "can't combine reasoning with tools on Chat Completions". A Chat Completions request that includes tools fails "even when you don't send reasoning_effort, because these models default to medium", and "an application that calls tools through Chat Completions can start failing after you upgrade its deployment from an earlier reasoning model." o1, o3, o3-mini and o4-mini all support Chat Completions with tools today, so this is exactly the path an auto-upgraded deployment takes. Microsoft's fixes: move tool calls to the Responses API (recommended), or set reasoning_effort to none on every Chat Completions request that sends tools.

Two other silent changes from the same page: GPT-5.6 models default reasoning.context to all_turns, so "expect higher token consumption on multi-turn conversations even when your code doesn't change" (set current_turn to keep the old behaviour), and the price per token changes, as the pricing section shows.

For rows with no replacement listed, there is nothing for an auto-upgrade to target. Microsoft does not say what a Standard codex-mini or claude-haiku-4-5 deployment does on November 15 beyond the general rule that retired models return 410 Gone. Plan for failed requests.

Provisioned deployments: nothing happens for you#

"Provisioned deployments are NOT auto-upgraded. Provisioned customers must manually migrate to the replacement model." The policy says the replacement becomes testable in provisioned regions "approximately 30 days before retirement", which for November 19 is now. You have two routes: an in-place migration, where Azure moves traffic over "a 20-30 minute window" while the deployment keeps its name, or a side-by-side migration to a new deployment. You need quota for the target model first. Batch deployments always use the side-by-side route.

Preview rows: gpt-chat-latest and gpt-image-1#

Preview models follow a different rule. When Microsoft retires one, "customers are force-upgraded to a replacement (a newer preview version or the GA model) or the model is retired with no replacement. There's no option to remain on a retiring preview model." Customers get at least 30 days notice either way.

For gpt-chat-latest, the schedule already lists newer preview versions: 2026-08-06 (retires January 12, 2027) and 2026-10-07 (retires May 18, 2027). It does not name either as the replacement for the November 17 versions, so we would not count on a force-upgrade. If your deployment is set to auto-update to default, the working-with-models page says preview deployments update "when a new preview version is available starting two weeks after the new preview version is released".

gpt-image-1 2025-04-15 retires this Friday with no replacement listed. The GA image models on the schedule are gpt-image-1-mini (retires April 7, 2027), gpt-image-1.5 (December 16, 2026), gpt-image-2 (October 21, 2027) and the two gpt-image-2.5 variants (September 9, 2027).

Replacements: Microsoft's, and ours where it names none#

Why those picks:

• gpt-5.6-sol and gpt-5.6-terra are GA (version 2026-07-09) with retirement dates of January 11, 2028. Microsoft's models page lists a 1,050,000-token context window (922,000 input, 128,000 output) and a February 2026 training cutoff, against 200,000 input and 100,000 output for the o-series and an October 2023 cutoff for o1. No access request is needed, but a quota request may be, depending on your subscription's quota tier (Tier 5 and Tier 6 have quota by default). Compare cutoffs on our knowledge cutoff table, and see our GPT-5.6 Sol, Terra and Luna guide for what separates the three.
• For o1-pro and o3-pro, the reasoning page documents a reasoning.mode of pro for GPT-5.6 models on the Responses API, for "difficult tasks that justify more model work". Microsoft does not tie it to the pro-model retirements, but it is the closest documented equivalent.
• codex-mini was Responses-API-only, and so are both suggestions. gpt-5.3-codex is GA until August 24, 2027 with a 400,000-token window; gpt-5.1-codex-mini is GA until May 15, 2027.
• claude-haiku-5-5 is GA on both hosting versions, per Microsoft's Claude models page, and its Azure-hosted version is on Data Zone Standard (US) like Haiku 4.5. It moves you from a 200K context and 64K output to 1M and 128K. Read our Haiku 5.5 guide first: Anthropic's migration guide says the same text produces about 30% more tokens on Haiku 5.5, which matters for its 100K-token price tier.

One oddity on the Claude side. Anthropic's own model deprecations page, which says its dates apply to Microsoft Foundry too, still lists claude-haiku-4-5-20251001 as Active with a tentative retirement "not sooner than October 15, 2026", and no deprecation notice or named replacement. Microsoft's schedule has a firm November 15 date for both Foundry versions. For Foundry deployments, plan against Microsoft's date.

Model pages on our side: o1, o1-pro, o3, o3-pro, o3-mini, o4-mini, GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.3 Codex, Claude Haiku 4.5 and Claude Haiku 5.5.

What it costs#

Azure OpenAI prices per million tokens, Global deployment, read from the Azure OpenAI pricing page on October 18, 2026:

The Azure page does not define where "short" ends; OpenAI's own pricing switches at 272K input tokens for both 5.6 models, per our first-party model data. Data Zone rates run about 10% higher (gpt-5.6-sol short context is $4.40 / $22.00), and Flex processing halves the 5.6 rates ($2.00 / $10.00 for Sol). The GPT-5.6 rows also list cache-write charges ($5.00 per million for Sol, $2.50 for Terra) that the o-series rows do not.

The Azure page has no rows for o1-pro, o3-pro or codex-mini. For reference only, OpenAI's first-party list prices in our model data are $150 / $600 for o1-pro and $20 / $80 for o3-pro. We have no first-party figure for codex-mini to show.

What that means for each move, at Global short-context list rates:

• o1 to gpt-5.6-sol: cheaper, about 73% less on input and 67% less on output.
• o3 to gpt-5.6-sol: 2x on input and 2.5x on output. Flex brings it back to $2.00 / $10.00 if your traffic can tolerate it.
• o3-mini or o4-mini to gpt-5.6-terra: about 1.8x on input and 2.7x on output. If that hurts, gpt-5.6-luna ($0.20 / $1.20) is on the same page, but it is our idea, not Microsoft's named replacement, so run evals.
• o1-pro to Sol pro mode: Microsoft says pro mode "bills those tokens at the model's standard rates" but uses more tokens than standard mode.

Claude on Foundry is billed in Claude Consumption Units, and Microsoft's CCU billing page says your cost is driven by "Anthropic's per-model token rates". Those first-party Anthropic rates, from our model data: claude-haiku-4-5 at $1.00 input and $5.00 output, claude-haiku-5-5 at $0.10 / $0.50 for prompts up to 100K tokens and $0.50 / $2.50 above. A new Haiku 5.5 deployment is billed in CCU automatically; older deployments stay on per-model token billing until you replace them.

The deployment change#

On Azure you call a deployment name, not a model id, so the migration is a new deployment plus a config change. Here is Microsoft's own example from the az cognitiveservices account deployment reference:

az cognitiveservices account deployment create -g yuanyang-test-sdk -n yytest-oai --deployment-name dpy --model-name ada --model-version "1" --model-format OpenAI --sku-capacity 1 --sku-name "Standard"

With the model swapped to the o1 replacement, your own resource names, and the Global Standard SKU code from Microsoft's deployment types page:

az cognitiveservices account deployment create \
  -g YOUR_RESOURCE_GROUP \
  -n YOUR_RESOURCE_NAME \
  --deployment-name gpt-5-6-sol \
  --model-name gpt-5.6-sol \
  --model-version "2026-07-09" \
  --model-format OpenAI \
  --sku-capacity 1 \
  --sku-name "GlobalStandard"

--sku-capacity is quota: "A value of 1 equals 1,000 Tokens per Minute (TPM)". Use gpt-5.6-terra with the same version for o3-mini and o4-mini. The version string 2026-07-09 is what the schedule and models page list; Microsoft's reasoning page labels the same models 2026-06-25 in its feature table, so if the create call rejects the version, check what the Models API endpoint above returns for your region.

Then point your code at the new deployment. Microsoft's Python sample from the Responses API page, after pip install openai azure-identity:

import os
from openai import OpenAI

# API key authentication
client = OpenAI(
    api_key=os.getenv("AZURE_OPENAI_API_KEY"),
    base_url="https://YOUR-RESOURCE-NAME.openai.azure.com/openai/v1/",
)
response = client.responses.create(
    model="MODEL_NAME",
    input="This is a test."
)
print(response.model_dump_json(indent=2))

The change for an o1 caller is the deployment name:

 response = client.responses.create(
-    model="o1",
+    model="gpt-5-6-sol",
     input="This is a test."
 )

If you are still on Chat Completions and send tools, either move that call to client.responses.create or add reasoning_effort="none" to it, per the warning above. The base URL is the same /openai/v1/ endpoint our Azure API compatibility reference lists.

For Provisioned deployments, Microsoft's in-place model family change is a PUT to the deployment with the new model name and version in the body. Its example on the working-with-models page moves a GlobalProvisionedManaged deployment to another model:

curl -X PUT https://management.azure.com/subscriptions/00000000-0000-0000-0000-000000000000/resourceGroups/resource-group-temp/providers/Microsoft.CognitiveServices/accounts/docs-openai-test-001/deployments/gpt-4o-ptu-deployment?api-version=2024-10-01 \
  -H "Content-Type: application/json" \
  -H 'Authorization: Bearer YOUR_AUTH_TOKEN' \
  -d '{"sku":{"name":"GlobalProvisionedManaged","capacity":100},"properties": {"model": {"format": "OpenAI","name": "gpt-4o-mini","version": "2024-07-18"}}}'

For an o3 PTU deployment you would set "name": "gpt-5.6-sol","version": "2026-07-09" and keep your own deployment path and capacity.

Claude Haiku 4.5 on Foundry#

Claude deployments use the Anthropic SDK against your Foundry resource. Microsoft's API key sample from Deploy and use Claude models, after pip install -U "anthropic", with only the deployment name changed and the sample's temperature and output_config effort lines removed (add back the settings you have tested on Haiku 5.5):

from anthropic import AnthropicFoundry

baseURL = "https://<resource-name>.services.ai.azure.com/anthropic" # Your base URL. Replace <resource-name> with your resource name
deploymentName = "claude-haiku-5-5" # Replace with your deployment name
apiKey = "YOUR_API_KEY" # Replace YOUR_API_KEY with your API key

# Create client with API key authentication
client = AnthropicFoundry(
    api_key=apiKey,
    base_url=baseURL
)

# Send request
message = client.messages.create(
    model=deploymentName,
    messages=[
        {"role": "user", "content": "What are 3 things to visit in Seattle?"}
    ],
    max_tokens=1048,
    thinking={"type":"adaptive"},
    stream=False
)

print(message.content)

Microsoft's Claude table lists "Extended thinking" for Haiku 4.5 and "Adaptive thinking" for Haiku 5.5, so a Haiku 4.5 request that sets a manual thinking budget needs that block rewritten. Our Claude on Foundry guide covers the deployment and auth side.

Checklist#

1. List your deployments on every Azure OpenAI and Foundry resource and note model, version, SKU and versionUpgradeOption for each. az cognitiveservices account deployment list -n NAME -g GROUP returns them.
2. Sort them by date: gpt-image-1 by Friday, codex-mini and claude-haiku-4-5 by November 15, gpt-chat-latest previews by November 17, the o-series by November 19.
3. Find your no-target rows. codex-mini, claude-haiku-4-5, gpt-chat-latest and gpt-image-1 have no listed replacement, so nothing will switch for you. Create the new deployment and move traffic.
4. Decide on auto-upgrade for Standard o-series deployments. Either let them upgrade to gpt-5.6-sol or gpt-5.6-terra on November 19, or migrate first on your own schedule. Do not leave one on NoAutoUpgrade by accident.
5. Fix Chat Completions tool calls before any o-series deployment becomes GPT-5.6: move them to the Responses API or send reasoning_effort="none".
6. Migrate Provisioned deployments by hand. Confirm quota and capacity for the target, then run an in-place or side-by-side migration.
7. Re-run evals and watch tokens. Check all_turns reasoning context on multi-turn traffic, the Haiku 5.5 tokenizer change, and cost against the table above.
8. Set a Service Health alert for Azure OpenAI Service, which Microsoft says is the service name to use, since there is no separate Foundry entry.

If you also run on AWS or Google Cloud, the Bedrock Claude 3.x shutdown lands two days earlier, on November 13, and the Vertex AI shutdowns are this week, on October 20 and 21.

FAQ#

When does o1 retire on Azure OpenAI?#

o1 version 2024-12-17 retires on November 19, 2026, along with o1-pro, o3, o3-pro, o3-deep-research, o3-mini and o4-mini. Microsoft names gpt-5.6-sol as the replacement for o1.

Will my o1 deployment be upgraded automatically?#

Standard-family deployments (Global Standard, Data Zone Standard, Standard) with the default upgrade setting are upgraded to the listed replacement at retirement. Deployments set to NoAutoUpgrade stop working. Provisioned deployments are never auto-upgraded.

What replaces codex-mini on Azure?#

Microsoft lists no replacement. gpt-5.3-codex and gpt-5.1-codex-mini are GA Codex models on the same schedule that also use the Responses API; that is our suggestion, not Microsoft's.

Is Claude Haiku 4.5 retiring on Microsoft Foundry?#

Yes. Microsoft's schedule lists both claude-haiku-4-5 versions (Anthropic-hosted and Azure-hosted) for retirement on November 15, 2026, with no replacement named. claude-haiku-5-5 is the GA Haiku model on Foundry.

What does a retired Azure OpenAI model return?#

Microsoft's lifecycle policy says all inference requests to a retired model return 410 Gone.

Continue Reading#

• AI Model Shutdown Dates 2026-2027: Calendar and Migration Plan - every provider's dates by month, including the December Foundry retirements.
• Bedrock Claude 3.x Shutdown on November 13: Which Claude Model to Move To - the AWS deadline two days before Foundry's first date.
• Vertex AI Shutdowns This Week: Gemini 2.5 on October 20, 16 Open Models on October 21 - the Google Cloud dates and replacement ids.
• Migrating Off Retired GPT Models in 2026: A Working Checklist - the eval-before-switch process for OpenAI's own API.
• GPT-5.6 Sol, Terra, and Luna: What Each Model Is For - the family Microsoft points the o-series at.

Sources#

All accessed October 18, 2026.

• Model retirement schedule, Microsoft Foundry - every date, version, lifecycle stage and replacement above, plus fine-tuned o4-mini dates
• Microsoft Foundry Models lifecycle and support policy - stage definitions, 410 Gone, auto-upgrade and Provisioned rules, preview force-upgrades, Models API terminology, no extensions
• Working with models - versionUpgradeOption values, capacity units, Provisioned in-place and side-by-side migrations
• Model versions in Microsoft Foundry Models - what happens at retirement per upgrade policy, partner model ids
• Azure OpenAI reasoning models - Chat Completions tool-calling failure on GPT-5.6, reasoning.context, reasoning.mode, feature tables
• Foundry Models sold by Azure - GPT-5.6 context windows, training cutoff, quota note
• Azure OpenAI pricing - per-token prices
• Understanding deployment types - SKU codes
• az cognitiveservices account deployment - CLI syntax and example
• Use the Azure OpenAI Responses API - Python sample
• Claude models in Microsoft Foundry - Haiku 4.5 and 5.5 availability, limits, deployment types
• Deploy and use Claude models in Microsoft Foundry - Python sample
• Claude Consumption Units billing - how Claude usage is priced on Foundry
• Anthropic model deprecations - Haiku 4.5 and Opus 4.5 status on Anthropic-operated platforms