Zhipu AI Open-Sources GLM-5.1 Model, Reportedly Outperforming GPT-5.4 on Coding Benchmarks
A 744B-parameter model outperforms top proprietary AI on coding tasks while costing nothing to run.
The AI landscape fractured in early April 2026 as two competing philosophies on model access were put on stark display. On one side, Anthropic confirmed its most capable model ever, Claude Mythos, but locked it behind a 50-company firewall called Project Glasswing. Priced at ~$25 per million input and $125 per million output tokens, its sole mandate is defensive cybersecurity for critical infrastructure partners. On the other, Chinese AI firm Zhipu AI released GLM-5.1 under a permissive MIT license, making it free for anyone to use, modify, and host.
GLM-5.1 is a technical powerhouse: a 744-billion-parameter Mixture-of-Experts (MoE) model that activates 40 billion parameters per forward pass, with a 200,000-token context window. Most notably, it reportedly outperforms both OpenAI's GPT-5.4 and Anthropic's Claude Opus 4.6 on SWE-Bench Pro, an expert-level software engineering benchmark. This performance, coupled with its open-source nature, presents a direct challenge to the closed, expensive approach exemplified by Mythos.
The week saw eight major model releases, including Google's open-weight Gemma 4 family and agentic models from Alibaba, but the GLM-5.1 vs. Claude Mythos dichotomy defined the era. The split is no longer purely about capability, as top-tier performance exists on both sides, but about control and accessibility. The industry's accelerating pace has outpaced consensus on who should wield its most powerful tools, forcing a clear choice between walled gardens for safety and open ecosystems for innovation.
- Zhipu AI's GLM-5.1, a 744B-parameter MoE model, outperforms GPT-5.4 on the SWE-Bench Pro coding benchmark.
- Released under a free MIT license, it starkly contrasts Anthropic's gated, $125/M output token Claude Mythos model.
- The releases highlight a major industry philosophical split: open accessibility versus controlled, expensive deployment for safety.
Why It Matters
Professionals now have a free, state-of-the-art coding AI alternative to expensive, closed APIs, reshaping tool accessibility and cost.