Open Source Models Score Low on ARC-AGI-2 Reasoning Benchmark (xcancel.com)

Recent evaluations of international AI models on the ARC-AGI-2 reasoning benchmark have revealed disappointing results, with performers like Kimi K2.5 scoring only 12% and others like Minimax M2.5 and GLM-5 achieving a mere 5%. This assessment, conducted by frontier labs and released on March 2, 2026, indicates that many state-of-the-art models are lagging significantly behind expectations for advanced reasoning capabilities—failing to reach the benchmarks set as far back as July 2025. These findings are significant for the AI and machine learning community as they underscore the challenges models face in achieving higher levels of reasoning and general intelligence. As the industry pushes towards creating more sophisticated AI systems, the low scores highlight the gap between current capabilities and the ambitious goals set for artificial general intelligence. The implications suggest a need for renewed research efforts and innovative approaches to enhance reasoning performance, particularly as competition intensifies in the pursuit of more effective AI solutions.