FrontierFinance Benchmark Assesses AI Agents for Investment Research

The deployment of AI agents in professional investment research is growing, yet existing benchmarks for financial AI primarily focus on narrow tasks like data extraction, which current models have largely mastered. These benchmarks often fail to capture the full complexity of an investor's workflow or adequately evaluate the open-ended, long-form answers required for real analyst queries. To address this, FrontierFinance has been introduced as a comprehensive and challenging benchmark. It comprises 220 expert-designed queries and over 11,500 source-attributed rubrics, covering six critical use cases across the entire investor workflow. The benchmark is designed to rigorously assess the "frontier intelligence" of AI models and agent systems using only publicly available data. Initial evaluations using FrontierFinance revealed that the tool harness (how the AI interacts with data and tools) plays a crucial role in determining quality and efficiency, often more so than the underlying model alone. Samaya's in-house system demonstrated leading performance, outperforming even the strongest frontier models like Claude Fable 5 at a significantly lower cost. Notably, the best open-weight model, Kimi K3, achieved performance close to proprietary models at a fraction of the cost. The benchmark also highlighted that "Screening & Discovery" and "Sector, Industry & Macro" remain the most difficult use cases for all systems.