Moonshot AI's K2.5 achieves frontier coding and agentic benchmarks at one-fifth the inference cost of leading closed models

Moonshot AI released K2.5 in January 2026, an open-source model that leads on open-source agentic and search benchmarks, matches closed-source coding performance, and achieves this at roughly one-fifth the cost per token of leading closed models. The source attributes the cost advantage to Mixture-of-Experts architecture, custom serving infrastructure, and in-house training optimization yielding roughly 2x token efficiency. Revenue doubled monthly following release, with customers including Cursor, Notion, Coinbase, and integration into Baidu, Alibaba, and Tencent products. This directly tests whether architectural efficiency can deliver frontier capability without frontier budgets and whether cost advantage is a structural wedge into enterprise AI spend currently facing accelerating consumption scaling.