Moonshot AI, an Alibaba-backed Chinese startup, released Kimi K2 Thinking — an open-source reasoning model designed to rival top closed systems like GPT-5 and Claude 4.5 Sonnet. The model demonstrated strong performance across multiple benchmarks while remaining dramatically cheaper to train. Official project: Kimi K2 Thinking.

• Benchmark results: Kimi K2 Thinking outperformed GPT-5 and Claude 4.5 Sonnet on several agentic and reasoning tests, scoring 44.9% on Humanity’s Last Exam — a new high.
• Progress: Major gains over its predecessor from four months ago, particularly in coding and reasoning, though still slightly behind the absolute frontier models.
• Capabilities: Can autonomously chain 200–300 tool calls for multi-step tasks and exhibits creative writing strengths.
• Cost efficiency: Estimated training cost is under $5 million — significantly below most current frontier model budgets.

Why it matters: With Nvidia’s Jensen Huang recently noting that China is “nanoseconds behind” in AI, Kimi K2 Thinking illustrates just how narrow that gap has become. By combining open-source transparency with frontier-level reasoning, Moonshot AI positions itself as a powerful challenger in both capability and cost efficiency.