Zhipu AI Open-Sources GLM-5.1 Model, Outperforming GPT-5.4 in Coding

In early April 2026, Zhipu AI released GLM-5.1, a massive 744-billion-parameter mixture-of-experts (MoE) model with 40 billion active parameters per forward pass and a 200K token context window. The model reportedly outperforms both Claude Opus 4.6 and GPT-5.4 on SWE-Bench Pro, an expert-level software engineering benchmark that tests real-world coding capabilities. Most significantly, Zhipu AI released GLM-5.1 under the permissive MIT license, making it completely free to use and self-host—users only pay for their own compute costs.

This open-source release came just hours after Anthropic announced Claude Mythos, its most capable model ever, would be locked behind a 50-company firewall called Project Glasswing. Mythos costs $25 per million input tokens and $125 per million output tokens, and is only available to select organizations like AWS, Apple, Microsoft, and major banks for defensive cybersecurity scanning. The simultaneous announcements highlight a growing philosophical split in AI development between open accessibility and controlled, security-focused deployment of the most powerful models.

The week saw eight major AI releases in seven days, including Google's Gemma 4 family, Alibaba's Qwen 3.6-Plus, and Microsoft's MAI models. However, the GLM-5.1 vs. Mythos contrast represents a fundamental industry divide: while some companies build increasingly capable models for public use, others restrict access to prevent potential weaponization. This tension between capability and control is becoming the defining narrative of advanced AI development as models grow more powerful.