Azure Content Understanding adds GPT-6.1 Sol as its default document model

Microsoft has widened the model choice in Azure Content Understanding to cover GPT-5.6, GPT-6 and GPT-6.1, and for most document extraction it advises starting with GPT-6.1 Sol, according to its announcement on the Foundry blog.

The post, published November 11, covers document, text, speech and video workloads. The new models sit alongside those the service already offered, such as the GPT-5 series, and Microsoft shares the evaluation data behind its updated advice.

Sol first, Luna for cost, Astra for expensive errors

Microsoft’s recommendation has three tiers. GPT-6.1 Sol is the balanced default for document extraction. GPT-6 Luna is the choice when token cost matters most. GPT-6 Astra is for teams that might pay a premium for a small quality gain. Microsoft describes Sol as close to the leader in its general availability (GA) accuracy tests, and as the leader in a separate preview extraction evaluation.

The numbers are close. In GA document extraction with only a schema, Astra scored 95.95% against Sol’s 95.74%. With five labeled examples the gap shrank to 0.11 percentage points. In the preview evaluation Sol came out ahead: 96.06% against 95.58% with a schema only, and 97.10% against 97.06% with examples. Microsoft says Astra costs about five times as much as Sol in completion tokens in these document tests.

Luna is the cheap option, and Microsoft’s data shows examples help it most. Five labeled examples lifted its extraction quality from 91.82% to 94.16%, while its token expense stayed well below Sol’s. The company cautions that examples add tokens and preparation work, and that quality has to hold on documents that were not used as examples.

Video and speech tell a different story

Microsoft repeats a point from its earlier GPT-5 guide: the best document model is not automatically the best at other jobs. Astra leads document segmentation in the GA benchmark, with a Segment F1 of 98.00% against Sol’s 97.10%. In video, among nine compared models, GPT-6 Sol has the highest headline Generation F1 (86.40%, plus or minus 0.70 points) and the best long-video segmentation mean (80.10%, plus or minus 1.20). GPT-6.1 Sol did best on short video ads, at 93.60% plus or minus 0.40, and its weakness is concentrated in long videos.

Speech results point to older models. GPT-5.2 has the highest GA field-classification mean and GPT-5.5 leads field generation. Microsoft notes these tasks measure structured fields pulled from audio, not transcription, and says it will update the post as it investigates whether the speech differences are significant. It also warns against merging the percentages into one ranking, since each task has its own scoring.

How to test the models on your own files

Microsoft’s practical advice is to pick two or three candidates and run the same analyzer, schema and labeled examples against each, changing only the model deployment. In Content Understanding Studio, users add a Foundry resource under Settings and configure default model deployments; Studio can deploy required models automatically if no suitable default exists.

For repeatable tests through the REST API, each analyze request can carry its own modelDeployments mapping, which overrides the resource defaults. The post’s example swaps the prebuilt-analyzer-completion deployment while keeping the embedding deployment fixed. Users should also check that each candidate model is available in the regions they need before testing; the post does not list regions.

Token price is only part of the bill

The analysis here is about cost. Microsoft’s cost figures are relative to Sol with a schema only (1.00x), use the evaluation’s own rates, and exclude service processing and other charges that do not change with the model. So the five-times figure for Astra describes completion tokens, not a full invoice, and the post sends readers to its billing explanation and current rates for budgeting. It publishes no per-token prices.

The model choice also sits next to another cost lever. In a separate post on November 9, Microsoft introduced an agentic mode, in preview, that runs a reasoning loop with tools for navigating tables, inspecting visuals and calculating. That post says the mode uses more tokens and takes longer than the default standard mode, and enables it with "workflow": "agentic" under API version 2026-06-01-preview. A team now has three dials to turn: model, labeled examples and extraction mode.

The figures are Microsoft’s own, measured on its own evaluation sets, and the source does not say whether outside parties have reproduced them. The company itself says outcomes vary widely by task. Microsoft is also promoting the work at its Ignite event, with a breakout session and a lab session on document workflows and agents.