· via dev.to (home feed)
Azure OpenAI retires o3 on November 19, 22 days before OpenAI's own API
Azure OpenAI retires the o3 model on November 19, 2026, while the same o3-2025-04-16 build on OpenAI's API runs until December 11. Teams tracking only OpenAI's deprecation page face an early breakage.

Same model, two retirement dates
According to a November 19 post by Versyka on dev.to, OpenAI's o3 reasoning model now has two different end-of-life dates depending on where it is hosted. On OpenAI's own API, the o3-2025-04-16 snapshot is scheduled to shut down on December 11, 2026. On Azure OpenAI, the deployment of that same version retires on November 19, 2026 — 22 days sooner.
The core problem, as the post describes it, is that most teams track the model maker's deprecation page, while the cloud hosting the model keeps its own retirement calendar with its own dates and its own suggested replacements. Azure does this, and so do Amazon Bedrock and Google Vertex AI. Sometimes the cloud's date falls later than the maker's; in this case it falls three weeks earlier, so anyone who planned a migration around OpenAI's December deadline has already lost most of their runway.
The wider retirement batch
The post lists the Azure OpenAI models retiring on November 19, 2026, along with their suggested replacements: o1 (2024-12-17) and o1-pro (2025-03-19) move to gpt-5.6-sol; o3 (2025-04-16), o3-pro (2025-06-10) and o3-deep-research (2025-06-26) also move to gpt-5.6-sol; and o3-mini (2025-01-31) and o4-mini (2025-04-16) move to gpt-5.6-terra. Taken together, the batch covers a large share of OpenAI's o-series reasoning models on Azure, from the original o1 through the o3 and o4 mini variants.
The deployment-name blind spot
A second trap the post flags is Azure's indirection layer. On Azure, application code does not call a model ID directly; it calls a deployment name that the team chose when the resource was created. If that deployment happens to be called something like reasoning-prod, a search of the codebase for "o3" will return nothing, even though o3 is exactly what is running and what is about to be switched off.
The practical consequence is that the audit has to happen in Azure itself, not only in the repository. The post recommends enumerating every deployment with its model and version — using az cognitiveservices account deployment list against the relevant resource group and account name, with a query that surfaces each deployment's name, underlying model and model version — and then comparing the results against Azure's model retirement page rather than OpenAI's.
Inventories need the host
The post's broader recommendation concerns model inventories. An entry that says "o3" is not enough; it needs to record the model, the version, the hosting platform and the deployment name, because the host determines the retirement date. The author notes that their own retirement calendar tracks cloud dates separately from the maker's, so a model running on Azure is checked against Azure's schedule while the same model elsewhere is checked against that host's.
Why it matters
Divergent retirement dates turn deprecation planning from a single lookup into a per-platform audit. A team that reads only OpenAI's deprecation page can believe it has until December 11, then find on November 19 that its Azure-hosted o3 deployments stopped responding — three weeks before the deadline it planned around. The same exposure exists for anyone serving models through Bedrock or Vertex AI. The fix is procedural rather than technical: inventory every deployment, record the host next to the model version, and treat each cloud's retirement calendar as the authoritative one for the models it runs.
- #openai
- #azure
- #model-deprecation
- #llm
- #cloud