The results look very different for other formats. For PDFs, Claude jumps to a 46 percent win rate. On Excel files (xlsx), it's 45 percent, and for PowerPoint presentations (pptx), 48 percent. In the "other" category—which includes various other formats—Claude Opus 4.1 matches human professionals at 50 percent.
GPT-5 shows the same trend: its scores for structured or visual formats like xlsx (36 percent) and "other" (49 percent) are much higher than for "pure text."
File format has a major impact on how well AI models stack up against humans in the GDPval benchmark. | Image: OpenAI
OpenAI doesn't give a reason for these differences. It's likely that human reviewers are influenced by things like layout and visual design. Models can score points with neat formatting, clear structure, and strong visuals in presentations, spreadsheets, or PDFs, even when the actual content isn't better. With plain text, that advantage disappears, and the AI has to compete directly with human-level writing and reasoning.
It's also worth noting that PDF, Excel, and "other" formats together make up over 80 percent of all deliverables. That helps explain why models like Claude Opus 4.1 end up with fairly high overall win rates, even though they still lag far behind humans on plain text tasks.
