Azure Content Understanding adds GPT-6, 6.1 and 5.6 models
Microsoft expanded the model families available to Azure Content Understanding and published evaluation data showing where its cheaper options come close to the premium one.
Azure Content Understanding models now include the GPT-5.6, GPT-6 and GPT-6.1 families, Microsoft said in a post on the Microsoft Foundry Blog published on November 11, 2026. The addition gives teams more options for document, text, speech and video workflows and sits alongside the GPT-5 series that was already supported. Microsoft also refreshed its model recommendations and published the evaluation data behind them.
The three starting points Microsoft suggests
For standard document extraction, Microsoft recommends starting with GPT-6.1 Sol, which it describes as close to the general availability accuracy leader and ahead on a separate preview extraction evaluation. Teams whose main constraint is token spend are pointed at GPT-6 Luna, described as much lower token spend with a quality trade-off to test. GPT-6 Astra is the premium choice, for cases where a small accuracy gain justifies the cost.
The headline numbers sit close together. On general availability document extraction with a schema only, Astra scored 95.95% against Sol's 95.74%. With five labeled examples the gap narrowed to 0.11 percentage points. Astra costs about five times Sol's completion tokens in those document tests.
The preview evaluation reverses the order: Sol scored 96.06% to Astra's 95.58% with a schema only, and 97.10% to 97.06% with five labeled examples. Microsoft's advice is to compare the two rather than assume the premium model wins.
Labeled examples as a cheaper dial
Content Understanding can run an analyzer from a schema alone, or with example documents stored in a knowledge base and labeled expected answers. In Content Understanding Studio, the Knowledge area is where those documents are connected and their labels reviewed or corrected.
In Microsoft's tests, adding five labeled examples lifted Luna from 91.82% to 94.16% on extraction while its token spend stayed much lower than Sol's. The post is explicit that savings are not automatic: examples add tokens and preparation cost, and the gain has to hold on documents that were not used as examples.
The best model changes with the task
Microsoft repeats a point from its earlier guide: the strongest document model is not automatically the strongest video or speech model.
Document segmentation favours Astra at 98.00% Segment F1 against Sol's 97.10%. For video, GPT-6 Sol posted the highest observed headline Generation F1 at 86.40% with a 0.70 percentage point margin, and the best long-video segmentation mean at 80.10% with a 1.20 point margin. GPT-6.1 Sol did best on short-video ads at 93.60% with a 0.40 point margin, but its weakness is long videos. On speech, GPT-5.2 had the highest general availability field-classification mean and GPT-5.5 led field generation; both measure structured fields from audio rather than transcription.
Each task uses its own scoring rules, so Microsoft warns against merging these percentages into one ranking.
How to compare candidates on your own files
There are two routes. In Content Understanding Studio you open Settings, add your Foundry resource and configure default model deployments; the Studio can deploy required models automatically when no suitable default exists. For repeatable testing, the analyze endpoint accepts a modelDeployments mapping on each request, and a request-level mapping overrides the resource defaults, so the analyzer and inputs stay fixed while only the completion deployment swaps.
Microsoft's suggested method: keep the same analyzer, schema, representative input set and labeled examples, change only the deployment, then compare quality, latency, token usage and failure rate.
Limits and open questions
The cost figures are relative to Sol with a schema only at 1.00x and deliberately exclude service processing and other charges that stay the same regardless of model, so they are not a billing estimate. The post gives no current per-token rates, no latency numbers and no region-by-region availability list, only advice to confirm each candidate exists in the regions you care about. It also says the speech findings will be revisited to see whether the differences are genuinely significant, which suggests that part of the data is the least settled.
What this means for building apps with AI
The transferable lesson is that model choice is a per-task setting rather than a default you pick once. A cheaper model plus five well-labeled examples beat a premium model on one document task in Microsoft's own tests, and the ranking flipped between two evaluations of the same family. Anyone wiring model calls into an application should plan to run the same inputs through two or three candidates and record cost and failure behaviour, not just accuracy. If you are building that kind of feature, Draftship's guide to adding AI to your app covers putting a model behind a working app.
Primary sources
- Content Understanding adds GPT-6.1, 6, and 5.6: Choose the right model for your workload · Microsoft Foundry Blog, Nov 11, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.