China's AI Upstart Moonshot Stuns Valley Again with a $4.6M Wonder (entropytown.com)

Moonshot AI released Kimi K2 Thinking, an open-weight reasoning model that the Beijing lab says matches or outperforms OpenAI’s GPT-5 and Anthropic’s Claude 4.5 Sonnet while reportedly costing just $4.6 million to train. Built on the July Kimi K2 release and a DeepSeek-derived architecture, Kimi K2 Thinking emphasizes heavy tool-use and autonomous planning—chaining 200–300 tool calls across browsers, spreadsheets and 3D software—and adds parameter scaling, optimizer tweaks and inference-cost controls. Observers report GPT-5-class scores on BrowseComp and Humanity’s Last Exam (e.g., 60.2% vs GPT-5’s 54.9% on BrowseComp; 44.9% vs GPT-5’s 41.7% on Humanity’s Last Exam), and the model continues Moonshot’s lineage of long-context strengths from prior Kimi releases. The $4.6M training figure, sourced to an anonymous insider, has been amplified in open-source communities. Why it matters: if the cost and performance claims hold, Kimi K2 Thinking demonstrates a major efficiency gap—achieving near-frontier results with open weights and relatively tiny training spend—driven by architecture reuse, data curation and cheaper Chinese compute. Backed heavily by Alibaba (36% stake) and now valued in the billions, Moonshot’s open, agentic model challenges Western incumbents’ moats and accelerates debate over how accessible frontier-capable systems will be. The release underscores that open-source labs can be competitive at the frontier and forces reassessment of assumed head starts held by large closed models.