Microsoft said on 11 November 2026 that 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 change gives teams more model choices inside Content Understanding, but the practical work remains model selection by workload rather than a single upgrade decision.
The new guidance keeps the same principle Microsoft used for the earlier GPT-5 series: compare quality, token use, latency, throughput, availability and regional deployment requirements on the same task. Content Understanding can use a schema alone, or it can use labelled example documents added through Knowledge in Content Understanding Studio.
Document extraction starts with GPT-6.1 Sol
For standard document extraction, Microsoft recommends GPT-6.1 Sol as the starting point. GPT-6 Luna is the lower-token option to test when token cost matters most. GPT-6 Astra is the candidate for cases where a small accuracy gain could justify a higher token bill.
In Microsoft’s general availability document extraction tests, Astra has a small lead over Sol for schema-only extraction, at 95.95% against 95.74%. With five labelled examples, the gap narrows to 0.11 percentage points. The same tests put Astra at about five times Sol’s completion-token expense.
That cost difference matters for high-volume document processing. Astra may still be worth testing where review costs, compliance checks or downstream decisions make each remaining error expensive. Microsoft’s own preview extraction evaluation also shows why the higher-priced model should not be assumed to win every task: Sol leads Astra there, at 96.06% against 95.58% for schema-only extraction and 97.10% against 97.06% with five labelled examples.
Labelled examples remain a separate control
Content Understanding can run an analyser with only a schema that describes the fields to return. Teams can also add example documents to the analyser’s knowledge base and label the expected answers. That gives teams another variable to test before moving to a larger model.
Microsoft’s tests show five labelled examples lifting Luna’s extraction quality from 91.82% to 94.16%. Its token expense remains much lower than Sol’s in those tests. The trade-off is that examples add preparation work and token use, and the quality still has to hold on documents that were not used as examples.
For teams running large document volumes, Luna with labelled examples is therefore a plausible test candidate. It should be compared against Sol and, where the error cost is high enough, Astra. The comparison only remains useful if the analyser, schema, input set and labelled examples stay the same across runs.
Segmentation, video and speech need separate tests
The document extraction recommendation does not carry across every Content Understanding task. For document segmentation, Astra leads Microsoft’s general availability benchmark with a Segment F1 of 98.00%, compared with Sol at 97.10%. Segmentation splits a file into logical documents, so it should be evaluated separately from field extraction.
Video results also vary by task shape. Among the nine models Microsoft compared, GPT-6 Sol has the highest observed headline video Generation F1 at 86.40%, with a margin of ±0.70 percentage points, and the highest long-video segmentation mean at 80.10%, with a margin of ±1.20 percentage points. GPT-6.1 Sol has the best short-video ads result at 93.60%, with a margin of ±0.40 percentage points, while Microsoft says its relative weakness is concentrated in long videos.
Speech has a different pattern again. GPT-5.2 has the highest general availability field-classification mean, while GPT-5.5 leads field generation. These speech results measure structured fields from audio, not transcription.
F1 balances finding the expected results with avoiding incorrect ones. Because each task has its own scoring rules, the figures should not be collapsed into one ranking of models. A model that leads one Content Understanding task may be the wrong economic choice for another.
Deployment testing should change one variable
Teams can test the new models in Content Understanding Studio by adding a Foundry resource in Settings and configuring default model deployments. Studio can deploy required models automatically when no suitable default exists. The same representative files should then be run across the candidate deployments.
For repeatable testing through the REST API, a request-level modelDeployments mapping can override the resource defaults. That allows the same analyser and inputs to be used while changing the completion deployment. Microsoft recommends checking the analyser’s supported models before testing, and confirming that each candidate is available in the required regions.
The cost figures in Microsoft’s model explorer compare model input and output charges using measured usage and evaluation rates, relative to Sol with schema-only extraction as 1.00×. They exclude service processing and other charges that do not change with model choice. Production budgets still need the current billing rates and the workload’s own usage profile.
Choosing between GPT-6.1 Sol, GPT-6 Luna and GPT-6 Astra for document, speech and video workloads is the kind of Azure AI Foundry decision covered in ECS's Core AI topic, alongside agentic AI, the EU AI Act and responsible AI. Sessions and practitioner discussions give attendees room to compare model selection, evaluation methods and production constraints beyond the benchmark figures. Bring representative workload questions and use the conference to test whether the cheaper, balanced or higher-accuracy option is the right starting point.
This article was written by the ECS editorial team with the assistance of Amal Agent, run.events' AI assistant, which supported the research, drafting and proofreading.