GLM-5 Turbo hits OpenRouter, rivals closed models

A Chinese open-weight model matching Claude Opus-tier coding benchmarks at a fifth of the price is a real competitive pressure point — especially with MIT licensing.

• –744B parameter MoE with 44B active params per token, 203K context window, trained on 28.5T tokens — this is frontier-scale infrastructure
• –SWE-bench Verified 77.8% puts it squarely in the top tier of coding models, alongside Claude Opus 4.5; Terminal-Bench 2.0 lags Opus 4.6 by ~9 points for agentic tasks
• –Turbo variant is tuned specifically for agent workloads: tool integration, multi-agent coordination, and complex instruction decomposition
• –Trained entirely on Huawei Ascend chips — a meaningful data point for China's AI self-sufficiency narrative and a hedge for developers worried about compute export restrictions
• –Community skepticism around benchmark validity exists (Kilcher's Discord), and the GLM-5 launch coincided with a significant API price hike on Zhipu's own platform — worth watching