Proprietary frontier artificial intelligence models are increasing their capability advantage over open-weight systems after a period in which the two categories had moved closer together, according to Citi.
The firm said the leading proprietary model’s score on the AA Intelligence Index increased to 58 from 53. The highest score among open models, meanwhile, increased to 46 from 44, taking the gap between the two groups to 12 points compared with 9 previously.
Citi said additional private evaluations have also indicated a wider separation in recent weeks. The difference has been particularly evident in long-horizon cybersecurity tasks and measures of reliability in production environments, according to the firm.
Citi said the previous convergence occurred partly during a period when regulatory considerations restricted the release of some advanced models. Open-weight developers also had access to model-distillation techniques and advanced offshore computing resources.
The firm said greater availability of compute could change that dynamic by enabling frontier laboratories to perform model distillation more efficiently than open-weight developers.
Recent model releases have also contributed to changes in measured capabilities. Citi cited Anthropic (NASDAQ:ANTP), which said on September 22 that Opus 5.5 raised the AA intelligence frontier by 9% while costing 40% less than Opus 5 for typical workloads.
Citi also noted that newer versions of Grok and Gemini have recorded increases in intelligence measures without corresponding increases in token pricing.
Looking at enterprise use of AI, Citi expects businesses to adopt combinations of different models to reduce dependence on a single system or provider. At the same time, the firm said AI developers could increasingly incorporate routing and orchestration functions directly within their model infrastructure.
Leave a Reply