Summary
Nick Saraev puts Moonshot’s newly released Kimi K3.1 Pro up against Anthropic’s Claude Opus 5.5 in a full end-to-end test. He begins with the benchmarks from Kimi’s landing page, then moves to practical builds.
On the published numbers, Kimi K3.1 Pro scores 65.7 on Terminal Bench 4.0 compared with 58.2 for GPT6 Asteris and close to Opus 5.5. It leads on SWE Marathon at 51.7, ahead of Opus 5.5 by almost five points, scores 61.2 versus 58.5 on Kimi Codebench 3.0 long horizon, and reaches 70 on Humanity’s Last Exam against 67.7 for Opus 5.5.
For the hands-on portion, both models received the same prompts, the same tools and their own harnesses at max effort, Kimi Code and Claude Code. Kimi averaged about 29 minutes 41 seconds per build against 47 minutes 2 seconds for Opus 5.5. Demos include a 3D mechanical keyboard configurator, a galaxy collision simulation, an SVG-only scroll-driven landing page, a tower demolition simulation and a Minecraft-style game. Saraev compares visual fidelity, performance and polish, and discusses how such a model can be used in real projects and businesses.
📺 Source: Nick Saraev · Published October 28, 2026
🏷️ Format: Comparison
