Moonshot's new open-source Kimi K2 Thinking has reset the frontier, outperforming proprietary models...
The AMW Read
Moonshot's K2 Thinking validates the Frame 2 debate by demonstrating that open-weight models can reset the frontier in agentic reasoning, fundamentally challenging proprietary dominance.
NoveltySignificance
Foundation Models · Player MapScaling Laws
Moonshot's new open-source Kimi K2 Thinking has reset the frontier, outperforming proprietary models in key agentic reasoning tasks. On the BrowseComp benchmark, K2 Thinking scored 60.2%, surpassing GPT-5's 54.9% and decisively beating Claude Sonnet 4.5. This high-performance open model, coupled with an API cost of just $0.15 per million input tokens, instantly democratizes top-tier agentic workflows. The LLM competition is now fundamentally a race for open source parity and real-world application cost dominance.
