Z.ai's MIT-licensed GLM-5.2 (753B MoE) competes with Claude Opus 4.8 on benchmarks. Opus leads in long-horizon SWE tasks, but GLM-5.2 wins on math and terminal agent benchmarks while costing up to 5.7x less.
On June 16, 2026, Z.ai released GLM-5.2, a 753B-parameter MoE model under MIT license. Compared to Anthropic's Claude Opus 4.8 (released three weeks earlier), GLM-5.2 achieves competitive performance on several benchmarks while costing up to 5.7x less.
Open-weight models are narrowing the gap with closed frontier models. GLM-5.2 is MIT-licensed with open weights, enabling self-hosting, fine-tuning, and quantization. Opus 4.8 is proprietary, accessible only via API with rate limits and deprecation risks.
GLM-5.2 demonstrates that open models can pressure closed models on price and openness. For agent workloads where output token cost dominates, GLM-5.2 is a cost-effective alternative. However, Opus still leads on long-horizon SWE agent performance.