Cheap AI Model Beats Frontier Model on Complex Task Using Recursive Query Method

Researchers rebuilt a Recursive Language Model (RLM) system using only OpenAI's budget-tier gpt-4.5-mini, testing it against a direct call to the more powerful frontier model gpt-5.6-sol. The task involved analyzing a 308,367-character context of 3,182 general-knowledge questions to identify the least-common answer category without any labels provided. The frontier model answered incorrectly, while the mini-only RLM arrived at the correct answer by breaking the problem into smaller chunks and processing them recursively through at least 238 model calls over roughly 102 minutes. Although the cheaper model made notable classification errors at the row level, it still reached the right final answer because the true minimum category had a large enough margin to absorb those mistakes. The experiment suggests that structured decomposition — rather than raw model capability — can be the deciding factor in complex analytical tasks.