Village study: Closed-source models have occupied the cost-effectiveness frontier, with open-source models optimal only in a 3% budget range

Village's LLM Atlas statistics show that open-weight models are currently the best choice in only 3% of the budget range, almost all concentrated at the lowest price points, occupied by Google Gemma 4 E4B and Xiaomi MiMo-V2.6-Flash; in June of this year this proportion was still 41%, with DeepSeek V4 Flash alone covering nearly one-third of the budget. OpenAI's GPT-6 Luna and GPT-6.1 Terra, as well as Anthropic's Claude Haiku 5.5, have reclaimed these ranges. The article offers three explanations: closed-source labs can use their own frontier models for knowledge distillation, cover various price points by model family, and achieve lower serving costs through larger-scale inference.