Open-weight Inkling-2-Small beats Claude Opus 5 and GPT-5.6 on speedrun benchmark

tessybarton · x · 2026-11-17

Inkling-2-Small from Thinkymachines outperformed Claude Opus 5 and GPT-5.6 Sol on the SuperTux task of SpeedRunBench, which measures whether agents can discover and refine a fast policy over repeated attempts using only a single score per attempt. Researchers say open-weight models are becoming increasingly competitive on interactive, long-horizon agentic tasks.

Related event: Open-Source Inkling-2-Small Tops Speedrun Bench, Beating Claude and GPT(2 posts)→

Original post →

More from Models

Models channel →