Azure Content Understanding adds GPT-5.6, GPT-6 and GPT-6.1
Azure Content Understanding now supports the GPT-5.6, GPT-6 and GPT-6.1 model families across document, text, speech and video workflows. Microsoft is also publishing evaluation data and recommendations to help teams balance quality, token cost and workload requirements.
- Azure Content Understanding now supports GPT-5.6, GPT-6 and GPT-6.1 model families.
- Microsoft recommends GPT-6.1 Sol as the starting point for standard document extraction.
- GPT-6 Luna can reduce token costs, while GPT-6 Astra targets workloads where accuracy justifies a premium.
- Model performance varies across document, text, speech and video tasks.
- Teams can compare deployments in Content Understanding Studio or through the REST API.
Microsoft recommends starting with GPT-6.1 Sol
For standard document extraction, Microsoft recommends starting with GPT-6.1 Sol. The company describes it as close to the general availability accuracy leader and says it leads the separate Preview extraction evaluation.
Teams that place the highest priority on token cost can test GPT-6 Luna, while GPT-6 Astra is aimed at workloads where a small quality improvement could justify a premium. Microsoft cautions that results vary widely by task, so the recommended models are starting points rather than universal rankings.
In the general availability document extraction evaluation, Astra recorded 95.95% for Schema only, compared with 95.74% for Sol. With 5 labeled examples, the difference narrowed to 0.11 percentage points. Microsoft says Astra’s completion token expense was about five times Sol’s in these document tests. In the Preview extraction evaluation, Sol led Astra, with 96.06% versus 95.58% for Schema only, and 97.10% versus 97.06% with 5 labeled examples.
Examples can change the cost and quality trade-off
Content Understanding can use a schema alone or draw on example documents stored in its knowledge base. Users can connect those documents in Content Understanding Studio and label the expected answers.
Microsoft’s tests found that five labeled examples improved Luna’s extraction quality from 91.82% to 94.16%, while its token expense remained much lower than Sol’s in those tests. That makes Luna with knowledge worth testing for high-volume processing, according to Microsoft. The company also warns that examples create token and preparation costs, and that their quality must hold on documents that were not used as examples.
Microsoft recommends keeping the analyzer, schema, representative inputs and labeled examples consistent when comparing models. Teams can change only the model deployment, then compare quality, latency, token usage and failure rate.
Different workflows need different model choices
The expanded model catalog does not produce one model that leads every task. Astra led the general availability document segmentation benchmark with a Segment F1 score of 98.00%, compared with Sol’s 97.10%. Segmentation divides a file into logical documents and should be tested separately from field extraction.
For video, GPT-6 Sol had the highest observed headline video Generation F1 among the nine compared models, at 86.40% ± 0.70 percentage points. It also had the highest Long-video segmentation mean, at 80.10% ± 1.20 percentage points. GPT-6.1 Sol produced the best result on Short-video ads, at 93.60% ± 0.40 percentage points. For speech, GPT-5.2 led the general availability field-classification mean, while GPT-5.5 led field generation.
Users can test deployments in Content Understanding Studio or through the REST API by passing a different modelDeployments mapping in each analyze request. Microsoft says teams should check supported models and regional availability before testing, then choose the combination of quality, cost, availability and throughput that fits their workload.
Microsoft Foundry: Content Understanding adds GPT-5.6, GPT-6, and GPT-6.1: Choose the right model for your workload
