Open-weights Kimi K2.6 tops Claude and GPT-5.5 in real-time coding contest

Moonshot AI’s open-weights Kimi K2.6 won Day 12 of Rohana Rezel’s AI Coding Contest, a sliding-tile word puzzle where models had to write TCP-connected game clients under a ten-second per-round wall-clock budget. Xiaomi’s API-only MiMo V2-Pro took second, GPT-5.5 third, and Claude Opus 4.7 fifth. Two Chinese labs occupied the top two slots; DeepSeek shipped malformed output and Nvidia’s Nemotron failed to compile.

The winning strategies diverged sharply. Kimi ran a greedy slide loop that oscillated inefficiently on small boards but kept generating moves on 30×30 grids where the scramble destroyed seed words and reconstruction was the only path to points. MiMo never slid at all, instead blasting every long word it could scan from the initial grid in a single packet — devastating when seeds survived, useless when they didn’t. Claude and Grok also declined to slide, which capped their ceiling on larger boards. Muse failed to filter short words against the scoring penalty and finished at −15,309, worse than not playing.

Rezel concedes the puzzle rewards aggressive claiming in ways safety-tuned models may resist, and that one contest doesn’t overturn broader benchmarks. Still, with Kimi K2.6 sitting at 54 on the Artificial Analysis Intelligence Index versus 60 for GPT-5.5 and 57 for Claude, the practical gap between frontier closed models and a model anyone can download locally is now small enough to surface in head-to-head results.