Today, with the release of Gemini 3 Flash, the bar for enterprise AI has shifted. We benchmarked Gemini 3 Flash against its predecessor, Gemini 2.5 Flash, on around 1,000 fields from our most difficult documents. The data reveals a 10 percentage-point improvement in accuracy, but the real story isn't just about general performance. It’s about complexity.
Gemini 3 Flash has emerged as a specialist in precision extraction, excelling specifically where other models lose focus. Unlike previous generations that primarily just processed text, Gemini 3 Flash applies deep contextual understanding to identify and capture critical data points from the most difficult enterprise layouts.
Precision where it matters most
In our latest testing on our updated evaluation sets, the model didn't just improve marginally. It unlocked entirely new categories of data for automated extraction.
- Long-document accuracy: Across our largest test set of lengthy documents, like contracts and agreements, Gemini 3 Flash achieved a 6-point lift over the previous version. This improvement confirms the model's ability to retain context and track information across multiple pages better than ever before.
- Lots of fields to extract: When tasked with populating extensive metadata templates, often requiring the extraction of dozens of distinct fields from a single file, Gemini 3 Flash achieved a 13 percentage-point increase over the previous version. While other models struggle to maintain accuracy as the number of requested fields grows, Gemini 3 Flash proves it can handle massive extraction tasks without losing focus or degrading data quality.
- PDFs: As the standard format for business operations, PDF performance is critical. Gemini 3 Flash achieved a 10 percentage-point lift on PDFs. This ensures reliable extraction from the most common enterprise file type, even when the document structure varies.
- Images: The model also showed significantly improved capabilities, registering a 9-point increase in accuracy on images compared to the previous version. By better reasoning over these files, Gemini 3 Flash can now extract data more reliably when the source material is imperfect, such as scanned or photographed documents.
Transforming knowledge work: Real-world use cases
The evolution of Gemini 3 Flash opens the door for enterprises to tackle more sophisticated, high-stakes use cases than ever before. With this level of accuracy, we’re finally seeing the barriers fall for automating complex workflows. Here’s how these improvements translate to tangible value in key use cases.
1. Financial services: Trusting the data in complex portfolios
In wealth management and insurance, teams often struggle to automate the ingestion of diverse portfolios that contain a mix of digital forms, scanned invoices, and rigid financial tables.
- The challenge: Extracting critical data from messy files (like tax documents and invoices) where information is crammed into dense layouts.
- The solution: With a 13-point improvement in extracting lots of fields, Gemini 3 Flash allows firms to automate the extraction of complex financial data with high confidence. This shifts the workflow from manual data entry to "review by exception," even for the most information-heavy mixed-media files.
2. Legal and compliance: Finding the needle in the haystack
Release windows are shrinking while expectations for compliance are soaring. Legal teams need to verify dates, clauses, and terms across voluminous contract bundles and regulatory filings.
- The challenge: Identifying specific termination dates, liability clauses, or varying terms buried deep within extensive agreement packages.
- The solution: With a 6-point lift on long documents, Gemini 3 Flash can ingest comprehensive contract bundles and identify relevant endpoints without losing the "thread" of the document. This ensures that critical dates and obligations are captured accurately, regardless of where they appear in the file.
3. Public sector and healthcare: Streamlining high-density intake
Agencies and healthcare providers process millions of documents annually (tax forms, medical intake records, inventory manifests, and more) that are characterized by extreme information density, often containing dozens of fields per page.
- The challenge: Most models suffer from attention degradation when faced with "wall-to-wall" data, hallucinating or missing fields when the density gets too high.
- The solution: Gemini 3 Flash’s high performance in PDF extraction showcases allows operations teams to fully automate the digitization of their most complex, data-heavy forms without breaking the massive forms into smaller chunks.
Try Gemini 3 Flash today
We encourage you to use Gemini 3 Flash to extract critical data points from your hardest content: scanned images, handwritten forms, and long contracts. Gemini 3 Flash is available today in Box AI.
