Citi said the capability gap between proprietary frontier artificial intelligence models and open-weight alternatives has begun widening following an earlier period of convergence.
According to the firm, the highest AA Intelligence Index score among proprietary models increased to 58 from 53, while the highest score for open models rose to 46 from 44. That expanded the gap between the two categories to 12 points from 9 points.
Citi said private evaluations indicate that leading open-weight systems have fallen further behind in recent weeks, particularly in long-horizon cybersecurity tasks and production reliability.
The firm attributed part of the previous narrowing in the capability gap to regulatory concerns that had constrained releases of advanced proprietary models. During that period, Citi said open-weight developers also benefited from model distillation and access to advanced offshore computing capacity.
Citi said increasing availability of computing resources could allow frontier AI laboratories to conduct model distillation more efficiently than their open-weight competitors, potentially affecting the relative capabilities of the two categories.
The firm also pointed to recent model releases as evidence of changes at the frontier. Citi noted that Opus 5.5 increased the AA intelligence frontier by 9% and costs 40% less than Opus 5 on typical workloads, based on information released by Anthropic (NASDAQ:ANTP) on September 22.
Citi added that successive releases of Grok and Gemini have increased intelligence scores while maintaining unchanged token pricing.
For enterprise adoption, Citi expects companies to use portfolios incorporating multiple AI models as a way of reducing reliance on individual systems. The firm also said AI laboratories could increasingly integrate model-routing and orchestration capabilities directly into the model layer.