Latest in AI
Azure's November 19 Model Retirements: Check Your Deployment Before Switching
Azure lists o-series retirements for November 19, but replacement availability differs for Batch, Regional Standard and provisioned deployments. Check the migration path.
If your Azure application still uses an o-series reasoning model, the important question today is not just which model Microsoft recommends next. It is whether that replacement exists for the way you buy and deploy inference.
Microsoft's schedule lists November 19, 2026 retirement dates for several o-series models. This article describes that published schedule and the documented migration constraints. It is not a report that I have observed every affected endpoint failing.
Identify the model behind the deployment name
The schedule lists o1, o1-pro, o3, o3-pro and o3-deep-research with GPT-5.6 Sol as the recommended replacement. It lists o3-mini and o4-mini with GPT-5.6 Terra. These recommendations have deployment-specific exceptions. Read the current retirement schedule.
I would inventory the Azure resource, deployment name, underlying model version, region and deployment type together. An application setting called “reasoning-production” does not tell you which model is serving it.
Also identify scheduled jobs and internal tools. A dashboard showing successful interactive requests does not establish that an overnight batch workload has a valid destination.
Read the replacement footnote before creating anything
Microsoft says GPT-5.6 Sol and Terra are not offered for Batch or Regional Standard deployments. Sol also has limited Global Provisioned coverage.
The schedule offers untested starting points where those replacements are unavailable: GPT-5.4 or GPT-5.4 Mini for Global Batch and US Data Zone Batch; GPT-4.1 or GPT-4.1 Mini for EU Data Zone Batch; and GPT-5.1 for Regional Standard where available. Provisioned customers must check regional availability. These are not tested equivalents. See footnote 4.
My first decision would therefore be whether to preserve the existing deployment boundary or deliberately change it. Do not silently move a region-bound workload to global processing just to make a recommended model selectable.
Separate automatic upgrades from manual migration
Microsoft's lifecycle policy describes rolling automatic upgrades for applicable Standard deployment types. Provisioned deployments are not automatically upgraded and need manual migration. The policy describes retired-model inference as returning 410 Gone. Review the lifecycle policy.
The migration guide adds that NoAutoUpgrade stops the deployment at retirement, while Batch follows a side-by-side migration path. Read the migration process.
Do not infer the actual state of your resource from the date alone. Check its configuration, Azure notices and observed responses. An automatic change can keep an endpoint responding without preserving the application's behavior.
I would assign an owner to each deployment rather than marking the whole subscription “handled.”
Check research tools separately from the base model
The classic Deep Research tool documentation ties its deprecation to the November 19 retirement of o3-deep-research. It recommends the latest Agents GA API, current GPT models and web search for a research experience. The same page still contains older instructions referencing o3-deep-research, so those examples should not be read as a fresh recommendation to deploy the retiring model. Read the classic tool notice.
For a research application, I would separately verify source retrieval, citations and completion behavior. A replacement model answering a short question does not demonstrate parity with a multi-step research workflow.
Compare the application contract and the bill
Microsoft's migration guidance warns about changes to parameters, tool definitions, structured outputs, calling APIs, latency and cost. It recommends assessing the target's region, deployment type and quota, not just its benchmark score. Review the assessment guidance.
My proposed acceptance sheet would contain one row for each behavior your product promises. Can it accept the longest real input? Does it return the fields your parser requires? Can it call the intended tools without adding an unintended action?
For cost, compare representative completed tasks rather than assuming a lower advertised token rate means a lower bill. Check the target's current input, output and caching prices, then include retries and reasoning usage where billed. I am not quoting a universal replacement price or claiming identical context limits.
Make the fallback usable after retirement
A rollback plan that depends on the retired model may no longer be a plan at all. I would define a supported fallback, a reduced-function mode or a clear unavailable response before increasing traffic.
Keep a record of the chosen replacement and the unresolved differences. Validate the exact deployed configuration, not a playground result from another region or deployment type.
These are proposed migration checks, not tests I have run on your Azure environment. The practical finish line is a supported workload with known behavior and cost, not merely a different model name in configuration.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help AI and infrastructure teams explain product changes through clear technical writing and practical storytelling.