Azure Content Understanding Adds GPT-6.1: How to Choose Models for Quality and Cost

What was announced​

Joe Filcik, a Principal Product Manager on Microsoft Foundry, published the post on November 11, 2026. Content Understanding now covers document, text, speech and video workloads with the new models, in addition to the GPT-5 series it already supported.

Microsoft's Learn documentation for service limits lists the supported chat-completion models. The list includes gpt-6.1-sol, gpt-6-sol, gpt-6-luna, gpt-6-astra, gpt-5.6-terra, gpt-5.6-sol and gpt-5.6-luna, alongside gpt-5.5, gpt-5.4 and gpt-5. The same table ties the newer models to the 2026-06-01-preview API version. The older gpt-5.x entries also list the 2025-11-01 GA version.

That detail matters for planning. Production teams on the GA API should confirm which model and API version pairings they can actually use before building a test plan.

Microsoft's starting points​

For standard document extraction, Microsoft recommends starting with GPT-6.1 Sol. The reasoning is that Sol is close to the accuracy leader in the GA evaluation and leads in the separate Preview extraction evaluation.

Microsoft stresses that these are starting points. Outcomes vary widely by task and content, so it tells customers to test on their own documents.

Sol versus Astra: the numbers disagree​

The document evaluations are the most useful part of the post, mainly because the two test setups give different answers.

• GA evaluation, schema only: Astra scored 95.95% and Sol 95.74%.
• GA evaluation, five labeled examples: the gap narrowed to 0.11 percentage points.
• Preview evaluation, schema only: Sol led, 96.06% to 95.58%.
• Preview evaluation, five labeled examples: Sol led narrowly, 97.10% to 97.06%.

Astra's completion-token expense was about five times Sol's in these document tests. A lead of a few tenths of a point is hard to justify at five times the token expense, unless Sol's remaining errors trigger costly manual review or wrong downstream decisions. Microsoft's advice is to compare the two models rather than assume the premium one wins. The GA and Preview results are separate evaluations, so they shouldn't be merged into one leaderboard.

Try labeled examples before paying for a bigger model​

Content Understanding can run an analyzer with only a field schema. Alternatively, you can add example documents to a knowledge base and label the expected answers. In Studio, the Knowledge area is where you connect and correct those labels.

Microsoft's data shows examples can matter more than model tier for the cheaper option. Five labeled examples raised Luna's extraction score from 91.82% to 94.16%, while Luna's token expense stayed much lower than Sol's. For high-volume pipelines, Luna with examples is therefore worth testing before moving up a tier.

The post is careful about the savings, though:

• Examples add tokens and preparation effort.
• Quality has to hold on documents that weren't used as examples.
• The savings aren't automatic.

Other workloads have different leaders​

Microsoft warns against reading the document winner as the best model for everything.

• Document segmentation: Astra led the GA benchmark with 98.00% Segment F1, against Sol's 97.10%. Segmentation splits a file into logical documents, so test it separately from field extraction.
• Video: among nine compared models, GPT-6 Sol had the highest headline Generation F1, at 86.40% ± 0.70 points. It also had the best long-video segmentation mean, at 80.10% ± 1.20 points. GPT-6.1 Sol did best on short-video ads, at 93.60% ± 0.40 points, and its weakness was concentrated in long videos.
• Speech: GPT-5.2 had the highest GA field-classification mean and GPT-5.5 led field generation. These tests measure structured fields pulled from audio, not transcription. Microsoft says it is still investigating whether the speech differences are significant and will update the post.

Caveats in the evaluation data​

An interactive comparison published alongside the post adds several qualifications:

• The video comparison covers nine models run in different execution cohorts. It is not a controlled model-only experiment and does not cover every historical model.
• The video error bars are one standard deviation, not confidence intervals, and they don't prove statistical significance.
• No token expense, cost or latency is paired with the video scores.
• Document token expense is relative to GPT-6.1 Sol with schema only, set at 1.00×. It is based on measured usage and the rates used in the evaluation. It excludes service processing and other charges that are the same whichever model you pick, and it is not a current price quote.
• F1 balances finding the expected results against avoiding incorrect ones. Generation rubrics use a 1–7 scale.

Because each task has its own scoring rules, the percentages shouldn't be combined into one ranking. They are Microsoft's own evaluations, which is useful but not independent. They are most valuable as a way to pick two or three candidates to try.

How to run your own comparison​

Microsoft's suggested approach is to hold everything constant except the model:

1. Use the same analyzer, schema, representative input set and labeled examples for every run.
2. Change only the model deployment.
3. Compare output quality, latency, token usage and failure rate.
4. Before testing, check the analyzer's supported models and confirm each candidate is available in your regions.

In Content Understanding Studio​

Open Settings, add your Foundry resource and configure its default model deployments. The Learn documentation describes the flow as selecting the Settings gear, choosing Add resource, picking the subscription, resource group and Foundry resource, then reviewing mappings and saving. There is an optional auto-deployment setting for required models when no default exists. Then select an analyzer, run it on the same files for each candidate, and review the extracted fields and raw response.

Through the REST API​

For repeatable testing, pass a different modelDeployments mapping in each analyze request. The mapping points the prebuilt-analyzer-completion alias at the candidate completion deployment. It points prebuilt-analyzer-embedding at an embedding deployment. A request-level mapping overrides resource defaults, so the analyzer and inputs stay unchanged while the completion model swaps. Run the same request once per candidate and compare the results and the response's usage data. That property reports token usage for the connected deployment along with Content Understanding's own meters.

Check support per analyzer first​

Learn notes that different analyzers support different models. To see what a given analyzer accepts, call GET /contentunderstanding/analyzers/{analyzerId}?api-version=2025-11-01 and read the supportedModels object. A model family announcement doesn't guarantee that every model works with every analyzer or region.

Billing and operations​

Content Understanding runs generative work on your own Foundry deployments. Learn's Foundry Models page notes that standard pay-as-you-go GPT-6 deployments use separate pricing categories for short-context and long-context requests. That is a further reason not to treat relative token figures as a budget. You are billed for tokens on the connected deployment, plus Content Understanding's own meters. Use Microsoft's pricing explainer and current rates for budgeting.

Bottom line​

This is a practical testing guide rather than a leaderboard. For most document pipelines, start with GPT-6.1 Sol, test Luna with labeled examples where volume makes token spend matter, and reserve Astra for cases where remaining errors are expensive. Evaluate segmentation, video and speech separately, because their leaders differ. Before committing, confirm that the model is supported by your analyzer, API version and region.