Microsoft has expanded Azure Content Understanding’s supported models to include the GPT-5.6, GPT-6 and GPT-6.1 families for eligible document, text, speech and video analyzer workflows. Its November 11, 2026 announcement also publishes model-selection guidance: start document extraction with GPT-6.1 Sol, compare GPT-6 Luna when token spending matters, and test GPT-6 Astra when a small quality improvement could justify a premium.
This is an Azure service-support update, not an announcement launching those GPT model families. The practical change is that eligible analyzers gain more deployment choices alongside existing models, including the GPT-5 series.
Microsoft’s own results give developers a useful reason to compare those choices carefully. Astra narrowly leads one document-extraction evaluation, while Sol leads another. For speech, older models still top some reported measurements.
API Support Is Broader Than Preview Alone
Microsoft Learn’s current release notes list the expanded model-family support in eligible workflows across both API versions:
| API version | Status | What developers should know |
|---|---|---|
2025-11-01 | Generally available | The release notes list the expanded model support for eligible analyzer workflows. |
2026-06-01-preview | Public preview | Also supports eligible workflows, alongside preview features with their own requirements. |
The new families therefore do not carry a universal requirement to use 2026-06-01-preview. Microsoft separately identifies features such as agentic document analysis as preview-API functionality.
That distinction matters for implementation planning. A team evaluating a different completion model should first check whether its existing GA analyzer supports it. There is no reason to introduce a preview dependency solely because a newer model appears in the catalog.
Microsoft warns that the preview API has no service-level agreement and is not recommended for production workloads. Model eligibility is a separate check: the release notes explicitly tell developers to inspect an analyzer’s supportedModels property before selecting a deployment. Catalog support is not blanket availability for every analyzer or region.
Sol, Luna and Astra: Where Microsoft Says to Start
Microsoft’s recommendations are most concrete for standard document extraction.
| Priority | Starting candidate | Basis in Microsoft’s evaluation |
|---|---|---|
| Balanced extraction quality and token spending | GPT-6.1 Sol | Close to the GA extraction leader and ahead in the separate preview evaluation. |
| Lower token spending | GPT-6 Luna | Lower measured token expense, with an extraction-quality tradeoff to validate. |
| Small accuracy gains where errors are expensive | GPT-6 Astra, compared with Sol | A narrow lead in GA extraction, accompanied by a substantial completion-token premium. |
These are starting candidates, not a promise that one model will suit every document collection. They also should not be read as an announcement that Microsoft has automatically changed every existing analyzer’s configured default.
Astra’s Lead Depends on the Evaluation
Microsoft reports the following extraction results:
| Evaluation configuration | GPT-6.1 Sol | GPT-6 Astra |
|---|---|---|
| GA evaluation, schema only | 95.74% | 95.95% |
| Preview evaluation, schema only | 96.06% | 95.58% |
| Preview evaluation, five labeled examples | 97.10% | 97.06% |
Astra’s schema-only lead in the GA evaluation is 0.21 percentage points. Microsoft says that gap narrows to 0.11 percentage points with five labeled examples. In the separate preview evaluation, Sol leads under both configurations.
These are Microsoft’s first-party measurements, not independent tests. Microsoft describes LLM matching as using an LLM to judge whether extracted field values match expected values. The percentages should therefore be read within that evaluation method, not treated as guaranteed accuracy on customer documents.
The reversal is the consequential detail. Paying for Astra does not consistently buy the highest reported extraction score, even within Microsoft’s published comparisons. I’d use Sol as the initial baseline and give Astra a targeted trial on fields where mistakes trigger costly review or downstream decisions.
How to Compare Deployments Without Changing the Workload
A useful comparison keeps the analyzer, input files, schema and labeled examples fixed while changing the completion deployment. Otherwise, a quality difference could come from several changes at once.
Check the Analyzer’s Supported Models
The release notes point developers to the analyzer’s supportedModels property. The opened SMF Clearinghouse technical walkthrough reproduces this GA request shape from Microsoft’s model-deployment documentation:
GET /contentunderstanding/analyzers/{analyzerId}?api-version=2025-11-01
Replace {analyzerId} with the analyzer you intend to use and inspect its returned model support. This is a relative request path, not a complete authenticated request.
Keep three names separate:
- Display name: GPT-6.1 Sol, the name used in the announcement.
- Model ID:
gpt-6.1-sol, identified in the walkthrough’s documentation analysis. - Deployment name: the name assigned to an actual deployment in your Azure setup.
The deployment mapping needs that last value. Pasting a display name into it does not identify the deployment you created.
Configure Studio With the Same Candidates
Microsoft’s workflow, also described in the opened implementation coverage, starts in Content Understanding Studio → Settings. Add your Foundry resource and configure its default model deployments.
The more detailed technical walkthrough describes selecting the subscription, resource group and Foundry resource, then reviewing and saving the mappings. Studio also offers automatic deployment of required models when no suitable default exists.
For a controlled comparison, reuse the same analyzer configuration and representative files as you switch candidates. If you add examples through Studio’s Knowledge interface, keep those examples and their corrected labels identical across each model run.
Override the Completion Deployment Per REST Request
The announcement’s REST approach uses modelDeployments to override resource defaults for an individual analysis request. Its example maps the completion alias to a candidate deployment while keeping the embedding deployment fixed.
A request body in that shape looks like this:
{
"inputs": [
{
"url": "<representative-input-url>"
}
],
"modelDeployments": {
"prebuilt-analyzer-completion": "<candidate-deployment-name>",
"prebuilt-analyzer-embedding": "<embedding-deployment-name>"
}
}
This body accompanies the analyzer’s :analyze request, using the appropriate API version. It illustrates the mapping, not a complete authenticated client.
For successive runs, change <candidate-deployment-name> to your Sol, Luna or Astra deployment name. Where the analyzer uses embeddings, leave <embedding-deployment-name> unchanged. The technical walkthrough reproduces this request pattern and explains the distinction between the completion alias and its deployment value.
Per-request overrides make this particularly useful for evaluation: the test selects its candidate explicitly without requiring a resource-default change for every run.
Record More Than an Overall Accuracy Score
Alongside extraction quality, record latency, token usage, failures and the fields that require human correction. Check deployment availability and throughput against the regions and processing volume your application needs.
Frequently Asked Questions
3 questions
1Do the New Models Require the Preview API?
No universal preview requirement is listed. Microsoft Learn lists eligible workflows in both 2025-11-01 and 2026-06-01-preview. Check the analyzer’s supportedModels property and the requirements of any additional features.
