10/09/2026
Building on Microsoft Foundry? Scheduled model retirements can have unpleasant consequences for your agentic solutions.
Microsoft lists gpt-4o version `2024-05-13` for retirement on 1 October 2026, with gpt-5.1 as the replacement. Depending on the deployment type and settings, Microsoft may upgrade the model automatically. Provisioned deployments need manual migration, while `NoAutoUpgrade` means the deployment stops working when the model retires.
The application may remain online while its behaviour changes. Response formats, tool calls, latency and cost can differ. Prompts, tool definitions and calling code may also need adjustment.
For a specialist solution, I recommend treating the replacement model as an application change. Run representative business tasks through it, including difficult cases and human approval paths. Compare the results against agreed acceptance criteria before choosing the cutover.
Copilot Studio needs a more precise check. Its Microsoft-managed primary models follow a separate lifecycle. A Prompt tool connected to your own Foundry deployment inherits that deployment's lifecycle, although the current integration excludes GPT-5 and later. A connection to an external Foundry agent introduces the dependencies managed by that agent's owner.
I am also investigating whether hybrid designs with self-hosted models could address re-emerging risks around data privacy and control with AI. This remains an evolving matter, and the operational responsibilities are substantial. I would not prescribe self-hosting to every customer.
The immediate action is to record the exact model version and upgrade setting, assign a migration owner, test the replacement and agree a fallback that remains usable after retirement.
Microsoft's lifecycle policy:
https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/model-retirements
