Kimi K3 is Moonshot's July 2026 open frontier flagship (2.8T MoE, 1M context, always-on reasoning) for long-horizon work—coding agents, research, and messy multi-step help. Our tests put it near the top of open assistants on helpfulness, instruction following, and comprehension, with strong tool follow-through. Empathy is capable but cooler than Claude or Grok. Creative writing pays the reasoning tax: default max thinking favors agent scratchpads over sustained voice, so our tests still prefer Claude Fable and the old GPT-4.5 peak for prose. Moonshot itself notes everyday UX still trails Fable 5 and GPT-5.6 Sol. Best open pick when you want an agent that finishes hard work; not the warmest writing partner. Scored at cost-efficient reasoning (not max showcase).