Azure Content Understanding now supports GPT-5.6, GPT-6 and GPT-6.1 families for document, text, speech and video workflows. Microsoft recommends GPT-6.1 Sol as a starting point for ordinary document extraction, GPT-6 Luna when token cost dominates and GPT-6 Astra where the remaining errors may justify a premium.
The published evaluations show why the choice is not a simple model ranking. Astra led one general-availability document extraction test by 95.95% to Sol’s 95.74%, but cost about five times as much in completion tokens. In a separate preview extraction test, Sol narrowly led Astra. Adding five labeled examples raised Luna’s extraction score from 91.82% to 94.16%, offering another lever before moving to a larger model.
Different models also led segmentation, video and speech tasks. Microsoft warns against combining those percentages because each test has distinct scoring rules. Region availability, latency, throughput and failure rates also remain part of a production decision.
Teams can switch deployments in Content Understanding Studio or pass a request-level model mapping through the REST API while keeping the analyzer, schema and input set fixed. The practical addition is broader model choice within one analysis service; Microsoft’s own data still calls for testing representative files rather than assuming the most expensive option will be best.