GLM-5

GLM-5 is the latest frontier model from Z.ai (formerly Zhipu AI), China's leading AI research lab. Released in early March 2026, it's a Mixture-of-Experts (MoE) architecture with 744 billion total parameters and 40 billion active parameters per forward pass — delivering frontier-level performance at dramatically lower inference cost than dense models of equivalent capability.

The benchmark numbers are striking: 77.8% on SWE-bench Verified (software engineering), placing it among the top models globally for coding tasks. It achieves this while using DeepSeek Sparse Attention (DSA), which significantly reduces both training and inference costs without sacrificing long-context fidelity up to a 200,000-token context window.

Key features: - 744B MoE architecture with only 40B active params per pass (efficient inference) - 200K token context window with strong long-document recall - 77.8% SWE-bench Verified — top-tier for real-world software engineering - Strong agentic task completion: web browsing, tool use, multi-step planning - Available via GLM Coding Plan subscriptions starting at ~/month - GLM-5.1 variant launched March 27 reaching 94.6% of Claude Opus 4.6 on coding evals

GLM-5 is a compelling choice for engineering teams looking for a highly capable coding and reasoning model with strong price-performance ratios. The GLM Coding Plan makes it accessible as a subscription service, and the API is available via Z.ai for teams needing programmatic access. Not currently open-source, but pricing is significantly below comparable closed models from OpenAI and Anthropic.