Microsoft ships two-tier Azure Content Understanding update: GA efficiency gains, preview reasoning tools

Microsoft has updated Azure Content Understanding, its document-extraction service, with two releases aimed at different stages of deployment. The refreshed CU 1.0 API is generally available for production workloads, focused on cost and reliability; the new CU 2.0 public preview adds lower-latency processing, structural improvements, and a more expensive reasoning mode for difficult extractions.

CU 1.0 GA refresh

The CU 1.0 GA refresh (API version 2025-11-01) now supports the GPT-5 model series, from GPT-5.5 down through GPT-5, in standard, mini, and nano variants. Microsoft describes this as a flexibility play: developers can weigh accuracy against latency or cost using the same analyzer patterns, and existing provisioned throughput commitments carry over. The refresh also merges extraction and grounding more tightly in the processing flow. Microsoft's internal evaluations report up to 28 percent lower average inference token usage and up to 3 percent better accuracy for GPT-4.1 and GPT-5.2, plus up to 14 percent improvement in confidence accuracy as measured by AUROC. Confidence scores tell automation pipelines whether to pass a result through, route it to human review, or apply extra validation.

CU 2.0 preview

The CU 2.0 preview (2026-06-01-preview) includes a broader set of changes. Synchronous Read and Layout APIs return structured results immediately, accepting binary content or URLs without temporary service-side storage — aimed at scenarios like grounding an agent mid-interaction or validating an identity document. Advanced Contextualization, which underpins several new features, uses labeled examples and document knowledge during extraction while keeping training data in the customer's Azure Storage account. Microsoft's internal tests show up to 3.5 percent accuracy improvement and 22 percent lower token usage for custom analyzers; for the five new prebuilt tax analyzers (covering forms including 1065, 1120-S, 8865, Schedule K-1, and Minnesota State M1), the claimed reduction reaches 99 percent, with some extraction scenarios said to require no LLM tokens at all.

Classification gains in-page splitting for cases where a physical page contains the end of one document and the start of another, with confidence signals attached to splitting and routing decisions. The Layout analyzer adds signature region detection and document metadata extraction. Semantic chunking in prebuilt-documentSearch builds retrieval units from document structure rather than fixed character counts — relevant for RAG pipelines, where chunk boundaries directly shape retrieval quality.

Agentic mode and caveats

Agentic mode is an iterative extraction workflow for documents whose evidence is distributed across sections, such as long legal agreements, financial filings, or insurance records. It identifies relevant content, extracts values, and refines output through additional reasoning passes. Microsoft is explicit that this increases latency and token consumption compared with standard extraction. The same release pushes token efficiency in one direction and accepts higher token costs in another, and Microsoft's own guidance is that customers should test agentic mode on representative documents and adopt it only when the quality gain justifies the added expense. All of the accuracy and cost figures are internal evaluations; the production economics will depend on the documents and workloads in question.