SotA ARC-AGI-2 Results with REPL Agents (www.symbolica.ai)

OpenAI's latest results from the ARC-AGI-2 benchmarking contest reveal significant advancements in reasoning capabilities, particularly with their GPT 5.2 (XHigh) model, which scored 53.33% across 120 problems. The ARC-AGI framework, initially launched in 2019 and upgraded with a second iteration in 2025, serves as a standard for assessing fluid intelligence in AI. The latest innovations focus on the integration of agentic workflows through Agentica, which enhances reasoning by allowing agents to interleave thought processes with execution, making performance improvements possible when tackling complex tasks. The Agentica framework utilizes a persistent Python REPL environment, enabling agents to maintain state, execute code, and dynamically adapt their strategies through recursive delegation. This interactivity has led to an impressive score of 85.28% for Opus 4.6, a significant leap over previous models, demonstrating the potential for agentic learning that can generalize across different reasoning domains. This development not only highlights the trajectory toward more sophisticated AI reasoning systems but also sets the stage for future explorations in autonomous agent design within the AI/ML community.