Enterprises rely on complex data that are critical to operations in finance, engineering, consulting, and beyond. Yet these files have long been difficult for AI to process accurately. Our Box AI Enterprise Eval shows that while Claude Sonnet 4.5 excels across the board, its transformative improvement lies in understanding multi-modal data and reasoning over structured and unstructured content together. This advancement unlocks new automation capabilities for some of the most manual, time-intensive enterprise workflows, particularly in professional services, hospitality, energy, retail, and public sector.
At Box, we're committed to providing customers with access to the best AI models to power critical business processes. Today, we're excited to announce that Anthropic's Claude Sonnet 4.5 will be available in Box AI. We evaluated the model on one of the most important enterprise tasks—metadata extraction—and the results reveal a significant performance leap from it’s predecessor.
Superior performance across the board
We evaluated Claude Sonnet 4.5 and Claude Sonnet 4 on our latest proprietary extraction dataset spanning over 40,000 fields across 1,500+ documents. The dataset included diverse industry use cases and document types—invoices, contracts, research papers, transaction files, and government identification—with varying lengths and modalities (text, image, and multi-modal content). This evaluation was designed to stress-test enterprise-critical capabilities: structured data extraction, complex reasoning over dense text, parsing unstructured formats, interpreting concise high-signal content, and handling multimodal inputs. The expanded dataset provides comprehensive insight into model performance across the full spectrum of enterprise document processing needs versus our prior data sets.
The results were clear: Claude Sonnet 4.5 achieved a 4.1 percentage point improvement in average accuracy over Claude Sonnet 4. This consistent, high-level performance was evident across nearly every industry dataset—including professional services, hospitality, energy, retail and public sector—establishing it as a powerful and reliable model for general-purpose extraction tasks across diverse business documents.
Mastering multi-modal data and reasoning across content with both structured and unstructured data
- The model demonstrated significantly improved ability to extract data from image files and multi-modal documents, with accuracy for image heavy documents climbing to 80% compared to 67% for Claude Sonnet 4—a +13 point gain. This unlocks the potential to reliably pull structured data from scans, photos, and documents that combine text and visual elements.
- For compact documents that include both tables and text content like receipts, passports, and invoices, accuracy jumped to 84.2%—a massive 17 percentage point improvement from 67.2%. This is a game-changer for automating workflows that rely on small but dense documents.
What this means for your business
These enhanced multi-modal capabilities translate into tangible benefits and streamline operations across your organization:
- Professional Services: Cut client onboarding time from days to hours by automating tax document, bank statement, and ID processing. Eliminate manual data entry on timesheets and expenses so teams focus on billable work. Accelerate deal cycles with faster processing of SOWs and contract amendments.
- Hospitality: Enable instant guest check-in with automated ID and booking confirmation processing. Reduce front-desk errors and bottlenecks. Convert handwritten feedback and loyalty forms into actionable insights that drive personalized service and repeat visits.
- Energy: Process field reports, maintenance logs, and safety docs from remote sites instantly—even from photos. Accelerate compliance and reduce audit risk with reliable extraction from permits and technical diagrams. Cut billing cycles with automated invoice and meter reading processing.
- Retail: Optimize cash flow with automated supplier invoice reconciliation. Speed up customer service through faster returns and warranty processing. Transform POS data, timecards, and feedback into real-time insights without manual entry—enabling rapid response to trends.
- Public Sector: Reduce permit and benefit application processing from weeks to days. Improve accuracy in tax processing and eligibility determinations while lightening staff workload. Handle complex documents—building permits with drawings, registrations with photos—reliably and completely.
Get started today
Claude Sonnet 4.5 provides a powerful new option for enterprises looking to automate critical processes with greater accuracy and reliability. Its improved performance on multi-modal data—especially on previously hard-to-process documents.
