Zhipu's open-source GLM 5.2 rivals top US models at one-fifth the cost

Chinese AI lab Zhipu has open-sourced GLM 5.2, a model that rivals top US systems (e.g., Anthropic's) on agentic benchmarks – testing real-world task planning, coding, and debugging without human steering – while costing roughly one-fifth as much. The new metric driving the industry is 'intelligence per dollar'. GLM 5.2 is joined by DeepSeek V4 and Alibaba's Qwen, which supports over 200 languages, challenging Western dominance. Enterprises like Box and Harvey are already adopting a strategy of routing easy tasks to cheap open models and reserving expensive closed models for high-stakes work. Additionally, OpenAI and Broadcom's custom Jalapeño chip cuts inference costs by half.

But the promise of sovereignty – especially for African institutions – is tempered by reality. Downloading model weights is not the same as deploying a working system. Significant engineering is required for compute infrastructure, fine-tuning on local data, building application layers, and establishing compliance. Price reduction alone unlocks value mainly for organizations with existing technical capacity. The real question isn't just affordability, but whether the surrounding conditions – talent, infrastructure, regulatory readiness – exist to turn a cheap model into a practical tool.