Log in
Azure Content Understanding now supports the GPT-5.6, GPT-6, and GPT-6.1 model families, giving you more choices for document, text, speech, and video workflows. This adds to the existing set of supported models like the GPT-5 series. To help customers understand how to use this expanded set of options, we’re updating our model recommendations and sharing the evaluation data behind them.
In our GPT-5 model-selection guide, we described model selection as a mixing board: adjust quality and cost for each workload. The new models add range to those controls. Not every workload needs to be set to maximum, so we’ll share data here to help you make the call.
Choose your starting settings
For standard document extraction, start with GPT-6.1 Sol. Add GPT-6 Luna when token cost matters most, and test GPT-6 Astra when a small quality improvement could justify a premium.
| Your priority | Starting point | Why |
|---|---|---|
| Balanced quality and token cost | GPT-6.1 Sol | Close to the GA accuracy leader; leads the separate Preview extraction evaluation. |
| Maximum accuracy, premium budget | GPT-6 Astra and Sol | Often similar quality to Sol, with a small GA extraction lead and a substantial token premium. |
| Low token cost | GPT-6 Luna | Much lower token spend, with a quality trade-off to test on your documents. |
It’s important to say that outcomes vary widely depending on the task. Pick some options based on our evaluations, then test these choices on your own task. The best balance varies with the content and the work you need done.
Default to Sol; test Astra where errors are expensive
Sol and Astra often deliver similar quality, with the advantage depending on the workload. Astra has a small lead in GA document extraction (95.95% vs. Sol’s 95.74% for Schema only), and the gap narrows with examples (0.11 percentage points for With 5 labeled examples).
LLM match means using an LLM to judge whether extracted field values match the expected values.
Astra comes at a substantial premium: about five times Sol’s completion-token expense in these document tests. That can be worth testing when Sol’s remaining errors lead to expensive review or incorrect downstream decisions.
For demanding documents, compare the two rather than assuming the premium model wins. In the Preview extraction evaluation, Sol leads Astra instead (96.06% vs 95.58% for Schema only; 97.10% vs 97.06% for With 5 labeled examples).
Try examples before turning up the model budget
Content Understanding can run an analyzer with just a schema describing the fields you want, without any example documents. Or add example documents to its knowledge base and label the expected answers.
In Content Understanding Studio, use Knowledge to connect those documents and review or correct their labels. Our tests compare Schema only and With 5 labeled examples.
That gives you another control on the mixing board before changing models. Adding five labeled examples improves Luna’s extraction quality (91.82% to 94.16%), while its token expense remains much lower than Sol’s in these tests.
For high-volume processing, that makes Luna with knowledge worth testing before paying for a larger model. Savings aren’t automatic: examples add tokens and preparation costs, and quality must hold up on documents you didn’t use as examples.
Give each task its own channel
The previous guide’s central lesson still applies: the strongest document model isn’t necessarily the strongest video or speech model. Keep the mixing-board settings separate when the job changes.
- Document segmentation: Astra leads the GA benchmark (98.00% Segment F1 vs Sol’s 97.10%). Segmentation splits a file into logical documents, so test it is separately from extracting fields.
- Video: Among the nine compared models, GPT-6 Sol has the highest observed headline video Generation F1 (86.40% ± 0.70 percentage points) and the highest Long-video segmentation mean (80.10% ± 1.20 percentage points). GPT-6.1 Sol shows the best results on Short-video ads scoring 93.60% ± 0.40 percentage points. Its relative weakness is concentrated in long videos.
- Speech: GPT-5.2 has the highest GA field-classification mean, while GPT-5.5 leads field generation. Those measure structured fields from audio, not transcription. We will update this post as we do additional investigation as to whether there is really significant differences between these speech tasks as those with video and documents.
F1 balances finding the expected results with avoiding incorrect ones. Each task has its own scoring rules, so don’t combine these percentages into a single model ranking.
Explore the data, not another wall of charts
Choose Document, Text, Speech, or Video, then a task. Hover, tap, or focus a model to inspect it; document extraction shows Schema only and With 5 labeled examples together.
Measurement note: Graphs show GA evaluation averages. Token expense compares model input/output charges using measured usage and the evaluation’s rates, relative to Sol with Schema only (1.00×); it excludes service processing and other charges that are the same independent of model choice.
Use the billing explanation and current rates for your workload budget. The explorer keeps the task-specific metrics and qualifications beside the data.
Loading the interactive chart. If it does not appear, use the link below.
&
So what should teams do next?
The expanded catalog gives every channel on the mixing board a wider range. The next step is to choose two or three candidates for your workload and test which combination of quality, cost, availability, and throughput deserves the final setting.
Keep the comparison simple:
- Start with Sol for document extraction; compare Luna when cost matters and Astra when remaining errors justify the experiment.
- Use the same analyzer, schema, representative input set, and labeled examples for every run.
- Change only the model deployment, then compare output quality, latency, token usage, and failure rate.
- Before testing, check the analyzer’s supported models and confirm that each candidate is available in the region(s) that you care about.
Start in Content Understanding Studio
- In Content Understanding Studio, open Settings, add your Foundry resource, and configure its default model deployments. Studio can deploy required models automatically when no suitable default exists.
- Follow the Content Understanding Studio quickstart to select an analyzer and run it on your own representative content.
- Test each candidate against the same files. Review the extracted fields and raw response, and record the quality, latency, and usage that matter to your workload.
Or compare deployments through the REST API
For a repeatable evaluation, pass a different modelDeployments mapping in each analyze request. A request-level mapping overrides the resource defaults, so you can keep the analyzer and inputs unchanged while swapping the completion deployment. For example, you can call:
POST /contentunderstanding/analyzers/myInvoice:analyze
{
 "inputs": [
   {
     "url": "<representative-input-url>"
   }
 ],
 "modelDeployments": {
   "prebuilt-analyzer-completion": "<candidate-deployment-name>",
   "prebuilt-analyzer-embedding": "<embedding-deployment-name>"
 }
}
Run the same request once per candidate, then compare the extracted results and the response’s usage data. Start with the balanced recommendation, add the lower-cost option, and include the best-quality model only when the remaining accuracy gap matters to the workflow.
Come learn more at Ignite
Join us at Ignite to learn more about building document workflows and trusted agents:
- Trusted agents need Document Intelligence & Content Understanding Breakout Session
- Build agentic document workflows with Content Understanding Lab Session
Get started
- Try Content Understanding in Studio or Microsoft Foundry.
- Analyze a document with agentic mode (preview).
- Configure model deployments for your analyzers to compare models on your own content.
The post Content Understanding adds GPT-5.6, GPT-6, and GPT-6.1: Choose the right model for your workload appeared first on Microsoft Foundry Blog.
