Unstructured Transform outperforms frontier models on table parsing

The limits of frontier models

Many developers assume that leading AI models, often called frontier models, are sufficient for reading and extracting data from documents. While these models are highly capable at recognizing general text on a page, they often struggle with the structural nuances of complex layouts. Recognizing characters is only one part of the task; maintaining the relationship between values in a grid is another.

Unstructured tested their new Transform model against GPT-6 Astra, Claude Opus 5.5, and Gemini 3.8 Flash using 224 real-world documents. The results showed that while frontier models extracted most of the text, they also invented a significant amount of spurious text. Specifically, these models had hallucination rates roughly 50% higher than those of Transform.

The test highlighted a trade-off between broad capabilities and specific accuracy. Because frontier models are built to handle a wide variety of tasks, they may not be the most efficient or reliable choice for document processing. For applications where accuracy is paramount, a dedicated parser can focus on the specific details that downstream systems require.

Accuracy in complex layouts

Complex documents often contain tables with merged cells, multi-level headers, and forms with dozens of paired fields. These structures are difficult for general AI because the layout dictates the meaning of the data. If a number is extracted correctly but assigned to the wrong column, the resulting data becomes useless for automated decision-making.

Unstructured Transform was designed to handle these specific challenges by focusing on how it identifies and processes different regions of a page. It aims to preserve the structural integrity of the document so that the output remains faithful to the original source. This is particularly important for documents like lab test results, insurance claims, or manufacturing parts lists.

The model also preserves structural context that helps AI agents navigate information. It links captions to specific figures and maintains heading levels throughout the text. This allows an automated system to locate a specific section or understand a chart by reading its associated description without human intervention.

A multi-step parsing approach

To achieve better results on complex tables, Unstructured utilized a principle of composability. This means breaking the parsing work into separate steps so that each step can focus on a specific task. By doing this, the system can apply different levels of attention depending on what it is looking at on the page.

The model uses different contexts for different types of information. Some questions require a close look at a small region to determine if it is a table, a form, or a paragraph. Other questions require a broader view of the entire page to understand the hierarchy of headings or which caption belongs to which figure.

Processing dense financial tables separately helps preserve the relationships between values. The larger and more complex a table is, the more important it is to give it focused attention. This approach ensures that merged headers and spanning cells do not lose their meaning when the data is converted into a structured format like JSON or Markdown.

Cost and performance gains

The cost difference between using a general frontier model and a dedicated parser is substantial. In the test conducted by Unstructured, processing 100,000 pages would cost $11,200 with GPT-6 Astra and $5,000 with Claude Opus 5.5. In contrast, the same task would cost only $1,500 using Transform.

While Gemini 3.8 Flash was cheaper than Transform, it received the lowest score for table extraction. This suggests that while a cheaper model might work for simple text, it may fail on the complex layouts that businesses actually need to process. Users often pay a premium for frontier models that provide slightly higher text recognition but lower table scores and more hallucinations.

Companies can use Transform to pull typed key-value pairs from documents like credit agreements or ACORD forms. It can also turn dense financial statements into structured grids or create plain-text descriptions for diagrams and photos. Because the model provides page and location information, users can trace every extracted figure back to its source in the original file.