Microsoft has set a hard deadline: as of September 13, 2026, any application calling Azure Computer Vision API versions 1.0, 2.0, 2.1, 3.0, or 3.1 will receive error responses instead of analysis results. The company wants all affected workloads moved to the newer Computer Vision 3.2 API or the Image Analysis 4.0 GA release, each with its own trade-offs. If you don’t act, your image-tagging, OCR, and object-detection pipelines will silently break.

What’s Actually Being Retired

The retirement doesn’t touch your Azure Computer Vision resource itself. It targets the API version your code requests. Even a freshly provisioned resource will reject calls that specify one of the now-deprecated versions in the endpoint path or api-version parameter. Microsoft’s documentation explicitly lists five affected versions:

  • Computer Vision API v1.0
  • Computer Vision API v2.0
  • Computer Vision API v2.1
  • Computer Vision API v3.0
  • Computer Vision API v3.1

After the cutoff, any request hitting a URL like https://{endpoint}/vision/v3.1/analyze will fail, regardless of how current the underlying service is. The outage is not gradual—these endpoints will stop working entirely.

The notice appeared in Microsoft’s official Azure Vision what’s-new page, and Azure Advisor will surface a retirement recommendation for some subscriptions. But the burden of discovery falls squarely on engineering teams. No automatic mitigation changes your code.

What the Deadline Means for You

If you run applications that have been calling the Azure Computer Vision service for years, chances are high that at least one of them pins an old API version. This can happen in obvious places—like a Python script that hard-codes /vision/v3.0/—and in hidden corners: a vendor-supplied desktop tool, a legacy Windows service, or a scheduled task launched from a server you rarely touch.

For developers, the immediate task is a code audit. Search your repositories, CI/CD pipelines, configuration files, and even compiled third-party binaries for any reference to the retiring version strings. Don’t assume that a recent resource name in the portal means everything is fine; the resource may be modern, but the client might still ask for an old API.

For IT administrators and architects, the challenge is less about finding a single app and more about inventorying every possible caller across your entire estate. That includes line-of-business apps maintained by external vendors, test automation, and internal tooling. A missing entry in your inventory could mean a production outage next September.

For business owners, the risk is purely operational. If your product relies on image analysis—whether for automatic photo tagging, document scanning, or retail shelf recognition—a failure here directly impacts customers. The fix is well-defined, but it still requires development and testing time. Starting early avoids a last-minute scramble.

How We Got Here

Azure’s Computer Vision API has been through multiple generations. Versions 1.0 through 3.1 formed the core of vision services for years, offering capabilities like image description, optical character recognition (OCR), and object detection. But Microsoft has been signaling a shift for a long time.

  • 2022–2023: Preview versions of 3.0 and 3.1 were deprecated, nudging users toward the generally available 3.2 API. The company also launched Image Analysis 4.0 in preview, powered by the Florence foundation model, with new features like dense captioning and product recognition.
  • March 2025: The Image Analysis 4.0 preview APIs were retired, forcing early adopters onto the GA version.
  • Now: The final curtain for the pre-3.2 stack is set for September 2026.

The newer APIs aren’t just a refresh; they bring quantifiable improvements. OCR accuracy is higher, language coverage has expanded dramatically (164 languages for print), and the programming model has evolved. Microsoft’s long-term investment is clearly in the Image Analysis family, which now underpins both the legacy “Computer Vision” label and the current SDK.

The 2026 retirement eliminates the last vestiges of the old architecture. After that date, the only supported GA path for classic Computer Vision workloads is version 3.2, while the modern route runs through the 2023-10-01 REST API used by the Image Analysis SDK.

What to Do Now

There’s no single command to upgrade. Migration demands a systematic approach, but you can break it into four phases.

Find Every Affected Call

Start by checking Azure Advisor under Reliability > Service Upgrade and Retirement. But don’t stop there. Advisor can miss dependencies that aren’t directly tied to a resource you monitor. Then, run searches across:

  • Source control for strings like /vision/v1.0/, /vision/v2.1/, or api-version=.
  • Application configuration, environment variables, and deployment templates.
  • Network logs or API gateways to catch runtime requests. A call from an unexpected machine often reveals a forgotten script.

Document each caller in a register: who owns it, what endpoint it hits, what operation it performs, and which SDK/package version it uses. Keep unknowns visible—mark items as “vendor confirmation needed” rather than guessing.

Choose Your Migration Target

Microsoft recommends Computer Vision 3.2 as the first option. It’s the least disruptive path when your app just needs to keep working without a redesign. The API contracts are similar enough that many workloads will need only a version-string change and regression testing.

Consider the Image Analysis 4.0 GA route (specifically the 2023-10-01 REST API) when you’re already planning to modernize the application or need capabilities that 3.2 lacks. But be aware: Image Analysis 4.0 itself retires on September 25, 2028. That’s a separate lifecycle event, not a reason to avoid the upgrade, but you must plan for it. For custom model analysis or image segmentation that isn’t yet GA, you may need a direct preview REST call. Use that only with a documented justification and a clear plan to move off preview.

Here’s a quick decision guide:
- Minimal change, low risk: Go to Computer Vision 3.2.
- Already rewriting the app: Target 2023-10-01 (Image Analysis GA).
- Need segmentation or custom models now: Use 2023-04-01-preview REST, but own the preview risk.

Test Beyond a 200 Response

Changing the API version isn’t plug-and-play. The response structure may differ—field names, coordinate formats, confidence values, error codes. Build a representative set of production images (various formats, resolutions, languages). Run the same inputs through the new endpoint, and compare every piece of data your app consumes. Test failure paths too: invalid images, timeouts, authentication errors. Make sure any downstream database or UI handles the new output correctly.

Plan the Cutover

Set a date. Deploy to a non-production environment first. Do a controlled rollout, monitor errors, and keep the old deployment artifact until you’re confident. Remember: after September 13, 2026, rolling back to the old API version won’t restore service; the endpoint will be dead.

What to Watch Next

This retirement is part of a broader cleanup. Microsoft is steadily sunsetting older Azure AI APIs, and the 2028 deadline for Image Analysis 4.0 means even this migration isn’t permanent. The takeaway is clear: treat API version as a managed dependency, not a set-it-and-forget-it detail. Keep an eye on Azure service health updates, and build your applications to gracefully handle version transitions. When the next notice arrives, you’ll already have a register of callers, a testing playbook, and a process to adapt—without the panic.