Citi flag: Open model approaching the front, closed source premium starting to loose

Citicorp believes that the capacity gap between open and closed source frontier models has narrowed significantly. In the latest assessment of Artificial Analysis, the leading closed source model Claude Opus 5 scored 63, the open weight model Kimi K3 reached 60, with a gap of only 3 points. Actual use is validating this catch. As at 1 September, the open weight model had accounted for 60.9 per cent of the Token volume of Vercel AI Gateway for almost three months; however, 29 per cent of Token was carried out in June with less than 4 per cent of the expenditure, while 95 per cent was still spent at four major closed source laboratories in the United States. This means that open-source models are taking over high-frequency, fault-tolerant tasks on a cost-benefit basis, while closed-source models continue to maintain complex and high-risk workflows. As enterprises adopt multi-model pathways, closed source laboratories must continuously prove performance premiums, otherwise Token prices and Māori rates will be under pressure and more value may flow to the reasoning infrastructure, model route and application layer. Kim Xian