# Azure Content Understanding adds GPT-5.6, GPT-6 and GPT-6.1 model options

published: 2026-11-11T19:05:47.571Z
topic: infrastructure
source: https://devblogs.microsoft.com/foundry/content-understanding-adds-gpt-5-6-gpt-6-and-gpt-6-1-choose-the-right-model-for-your-workload/
url: https://aiagentnews.io/stories/azure-content-understanding-adds-gpt-5-6-gpt-6-and-gpt-6-1-model-options

Microsoft's Azure Content Understanding now supports the GPT-5.6, GPT-6 and GPT-6.1 model families for document, text, speech and video workflows. The Microsoft Foundry team also published evaluation data and starting recommendations for choosing among them.

For standard document extraction, Microsoft recommends starting with GPT-6.1 Sol, adding GPT-6 Luna when token cost matters most, and testing GPT-6 Astra when a small quality gain could justify a higher price. In the GA document tests, Astra scored 95.95% against Sol's 95.74% with schema only, but it used about five times Sol's completion-token expense.

Adding five labeled examples raised Luna's extraction quality from 91.82% to 94.16%, while its token spend stayed far below Sol's. Results differ by task: Astra leads document segmentation at 98.00% Segment F1 against Sol's 97.10%, and GPT-6 Sol posts the highest headline video Generation F1 among the nine compared models.

Microsoft advises running the same analyzer, schema and input set against each candidate model, changing only the model deployment. Developers can override the completion deployment per request through the REST API using a modelDeployments mapping, and should confirm each model is available in their region.

## Why it matters if you build agents

Agent builders working with documents or video can now pick a model tier per task instead of defaulting to the largest option.

Scores from one task, such as extraction, should not be used to rank models for segmentation, video or speech, since Microsoft notes each task has its own scoring rules.

Token cost and accuracy trade-offs should be measured on your own documents, because the published results are averages from Microsoft's evaluations.