GLM-5.3 Takes Third in Terminal-Bench 4.0

GLM-5.3’s performance makes open-weight models credible contenders for real terminal-based coding work, but the leaderboard is measuring the model-agent combination—not the model in isolation.

• –GLM-5.3 reached 41.82%, versus 37.27% for GPT-5.6 Sol.
• –Terminal-Bench 4.0 covers 66 tasks with five trials and up to eight hours per task.
• –The benchmark changed substantially, fixing 19 tasks and removing eight, so scores are not directly comparable with earlier versions.
• –Claude Code’s harness and maximum reasoning setting materially influence the result.
• –Full weights make the result more actionable for developers, though practical local deployment remains hardware-intensive.