Contact Us 1-800-596-4880
  1. Homepage
  2. IDP

  3. Supported Models

Supported Models

Einstein supports these predictive multimodal models:

  • OpenAI’s GPT-4o (gpt-4o) LLM

  • OpenAI’s GPT-4o Mini (gpt-4o-mini) LLM

  • OpenAI’s GPT-5.1 LLM

  • OpenAI’s GPT-5.2 LLM

  • Google’s GEMINI-2.0 Flash 001 LLM

  • Google’s GEMINI-2.5 Flash LLM

  • Google’s GEMINI-3.5 Flash LLM

  • Anthropic’s Claude Sonnet 4.6 LLM

  • Anthropic’s Claude Opus 4.6 LLM

Einstein GEMINI Flash 2.5 retires October 16, 2026

To prevent service interruption, migrate actions to a supported model by October 16, 2026.

Einstein OpenAI GPT-4o and Einstein OpenAI GPT-4o Mini retire January 31, 2027.

To prevent service interruption, migrate actions to a supported model by January 31, 2027.

When creating a document action, you can select the model and extraction settings such as PII masking or image recognition. Each model performs differently under various conditions, so select the one that aligns with your specific requirements.

For additional details on GPT models' lifecycle and deprecation, see Microsoft Foundry Models lifecycle and support policy.

This table compares the supported models by stability, image recognition performance, location callout support, and how each model provides confidence scores. The Notes column includes recommendations and limits for each model, such as supported languages and the maximum number of fields.

Model Name Description Stability Image Recognition Location Callouts Confidence Score Notes

Einstein OpenAI GPT-4o

Suitable for most tasks, performing well on documents in non-Latin languages. Can compare font sizes and identify certain font styles.

High

  • Standard performance

  • Low accuracy on reading checkboxes in forms

Not supported

Calculated

  • Ask the model to not make up information to prevent hallucinations

  • List of languages supported by OpenAI models

Einstein OpenAI GPT-4o Mini

Fast and useful for focused tasks. Tends to exhibit lazy reasoning.

High

  • Standard performance

  • Low accuracy on reading checkboxes in forms

Not supported

Calculated

  • Ask the model to not make up information to prevent hallucinations

  • List of languages supported by OpenAI models

Einstein OpenAI GPT-5.1

Suitable for most tasks. Improved instruction following and accuracy compared to GPT-4o.

High

  • Medium accuracy on reading checkboxes in forms

Not supported

Calculated

  • Meets EU data residency requirements

  • Ask the model to not make up information to prevent hallucinations

  • List of languages supported by OpenAI models

Einstein OpenAI GPT-5.2

Suitable for documents with charts, tables, and complex layouts due to improved visual understanding.

High

  • Medium accuracy on reading checkboxes in forms

  • Performs slightly better than earlier GPT models when reading checkboxes and tables

Not supported

Calculated

Einstein GEMINI-2.0 Flash 001

Good for analyzing images due to increased accuracy in this type of documents.

Standard

  • Medium accuracy on reading checkboxes in forms when processing one page at a time

Supported

Calculated

Einstein GEMINI-2.5 Flash

Good for analyzing images due to increased accuracy in this type of document. Performs quicker and with higher accuracy than Einstein GEMINI-2.0 Flash 001.

Standard

  • Medium accuracy on reading checkboxes in forms when processing one page at a time

Supported

Calculated

Einstein GEMINI-3.5 Flash

Fast and the strongest all-round performer in current testing. Handles structured documents and invoices, including non-English formats (localized dates, currency, and thousand separators). Leads in checkboxes and handwriting.

Standard

  • Best of the tested models at reading checkboxes and handwriting in multimodal mode

Supported

Self-reported

  • Ask the model to not make up information to prevent hallucinations

  • List of languages supported by Gemini models

Einstein Anthropic Claude Opus 4.6

High-accuracy extraction on structured and semi-structured documents with complex or varied layouts, including invoices. Favors conservative, no-guess answers over made-up values.

Standard

  • Performs well on handwritten documents

  • Medium accuracy on reading checkboxes in forms

Not supported

Self-reported

  • Supports up to 15 fields per document action

  • Ask the model to not make up information to prevent hallucinations

  • List of languages supported by Anthropic models

Einstein Anthropic Claude Sonnet 4.6

Good for general document extraction tasks requiring a balance of accuracy and speed. Performs well on structured and semi-structured documents with complex or varied layouts.

High

  • Performs better than Einstein OpenAI GPT-5.2 on handwritten documents

  • Medium accuracy on reading checkboxes in forms

Not supported

Self-reported

  • Supports up to 15 fields per document action

  • Ask the model to not make up information to prevent hallucinations

  • List of languages supported by Anthropic models

Confidence Scores by Model

Each model returns a confidence score for every extracted value. Depending on the model, the score is calculated or self-reported. This applies to all document action types.

  • Calculated: IDP derives the confidence score from model probabilities rather than self-reported estimates. This score reflects model certainty and is generally more stable, but it doesn’t guarantee that the extracted value is accurate.

  • Self-reported: IDP prompts the model to rate its own confidence when model probabilities aren’t available. This score reflects the model’s self-assessment and is generally less stable than a calculated score, but it provides a baseline estimate.

Self-reported confidence scores become less reliable as document processing complexity increases. Reliability depends not only on schema complexity, such as processing numerous fields or complex tables, but also on document quality, prompt logic, and structural variability across document groups. Low-quality scans, prompts that require data transformation alongside extraction, and inconsistent field availability across uploads can all degrade model self-assessment. Fields that are correctly returned as not found can show a low or 0% confidence score, which sends the document to human review even though the extraction is correct.

IDP continues to refine how it requests self-reported confidence scores, which reduces how often this happens. However, because the model estimates its own confidence, no level of complexity guarantees consistent scores. Prompt instructions about confidence, such as Fields that return "Not found" must have a 99% confidence rate, don’t reliably change the reported score.

To get more reliable results with self-reported confidence scores:

  • Keep the document action as simple as possible. Remove fields, tables, and prompts that aren’t required, and consider splitting large extractions into multiple document actions.

  • Test the document action against a large, representative set of documents, including documents where optional fields are missing, before moving to production.

  • Account for this behavior when setting confidence thresholds and assigning reviewers.

  • If confidence scores for not-found fields are critical for your use case, select a model that uses calculated confidence scores.

Model Errors

In some scenarios, the LLMs stop generating a response and provide an error code to understand the reason that caused this behavior. If the error occurs during the execution of a document action, IDP provides this error in the statusMessage attribute of the response object.

These are some of the most common error codes and their suggested fixes:

Error Code Error Description Suggested Fix

length

The request exceeded the maximum input token limit.

See Using long context under GPT-4.1 prompting best practices to understand how to refine your prompts.

MAX_TOKENS

The response exceeded the maximum output token limit.

Split the document or select an OpenAI model for this extraction.

SPII

The document potentially contains Sensitive Personally Identifiable Information (SPII).

Delete the conflicting field, or use GPT models to abstract the information.

For a complete list of error codes, refer to each of the model provider’s documentation:

See Also