July 2026 Summaries

Chartography is a sophisticated benchmark designed to evaluate the ability of AI models to interpret complex charts and graphical data, essential in professional fields such as medicine, engineering, and finance. Unlike traditional benchmarks that focus on simpler chart types like bar or pie charts, Chartography includes domain-specific graphics such as Sankey diagrams, Kaplan-Meier curves, and wind roses, requiring models to perform tasks like estimating unlabeled values and interpreting complex geometries. Despite advancements in AI performance on simpler chart reading tasks, current frontier models achieve only up to 45% accuracy on Chartography, highlighting challenges in visual reasoning, multi-step inference, and adherence to domain-specific conventions. The benchmark underscores the necessity for AI models to develop professional-level chart reading skills, which involve intricate visual interpretation and judgment comparable to human experts. Chartography thus serves as a critical tool for measuring AI's progress in understanding professional-grade visual data and for distinguishing between academic proficiency and practical competence in AI systems.

GDP.pdf is a benchmark designed to evaluate AI models' performance on tasks that reflect real-world professional workflows across 10 domains, such as medicine, law, and finance. With contributions from professionals who create and assess these tasks based on their own work experiences, the benchmark aims to highlight the gap in AI models' ability to handle routine tasks accurately. Notably, OpenAI's GPT-5.6 model, despite excelling in several other benchmarks, scored only 30.7% on GDP.pdf, indicating significant room for improvement in handling complex, context-dependent tasks typically performed by skilled professionals. The benchmark is publicly available for evaluation, with detailed results and methodology accessible online. The initiative underscores the rising importance of reliable AI performance in professional environments and has been cited by various organizations in their AI research and development efforts.