Vals AI Releases Finance Agent v2 Benchmark; GPT-5.5 Scores 51.76%, All Models Fall Below 40% Under Strict Grading

According to Beating, AI evaluation firm Vals AI released its second-generation Finance Agent v2 benchmark on May 14, testing financial analysis workflows through 927 expert-reviewed questions. GPT-5.5 topped the rankings with a 51.76% accuracy rate, closely followed by Claude Opus 4.7 (51.51%) and Claude Sonnet 4.6 (51.03%). The test required models to independently locate relevant sections across hundreds of pages of 10-K and 10-Q financial statements and complete multi-step calculations with precise intermediate figures.

Under strict grading standards requiring completely correct answers, all leading models' accuracy rates fell below 40%, with the hardest categories—financial modeling and precedent analysis—reaching only 23% at best. Among other models, Kimi K2.6 ranked fifth at 44.87%, followed by GLM 5.1 (44.79%) and DeepSeek V4 (44.08%). Compared to the previous version where Opus 4.7 scored 64.4%, the significant decline underscores that while AI handles simple retrieval, it remains far from replacing human analysts in finance's complex domain requiring strict numerical precision.