For makers and artists relying on open-source tools, the release of GLM-5.2 from Z.ai marks a significant shift in the landscape of text-only generative models. Originally restricted to coding plan subscribers on 13 June, the full weights were opened under an MIT license on 16 June. This 753-billion parameter beast, weighing in at 1.51TB, utilises 40 active parameters in a mixture-of-experts architecture. While Z.ai maintains a separate, non-open vision family including GLM-5V-Turbo, GLM-5.2 is strictly text input. It boasts a context window of one million tokens, a substantial jump from the 200,000 tokens supported by its predecessor.
The reception has been immediate and robust. According to Artificial Analysis, which runs the highly respected Intelligence Index, GLM-5.2 has overtaken the competition to become the leading open-weights model. The benchmark scores it at 51, edging out MiniMax-M3 and DeepSeek V4 Pro, both sitting at 44, and Kimi K2.6 at 43.
GLM-5.2 is the leading open weights model on the Intelligence Index v4.1. At 51, it leads MiniMax-M3 (44), DeepSeek V4 Pro (max, 44) and Kimi K2.6 (43)
However, this performance comes with a caveat regarding efficiency. The model is notably token-hungry, consuming 43,000 output tokens per task. This is a sharp increase from GLM-5.1’s 26,000 tokens and significantly higher than rivals like MiniMax-M3 (24k) and DeepSeek V4 Pro (37k).
Despite the verbosity, the model has secured second place on the Code Arena WebDev leaderboard, trailing only Claude Fable 5. This ranking is particularly noteworthy given that the leaderboard evaluates front-end web development and agentic coding workflows. It is impressive to see such high performance from a text-only architecture, challenging the assumption that visual input is essential for top-tier frontend coding capabilities.
Testing the model via OpenRouter, which aggregates it from nine different providers, reveals steep costs. Input is priced at $1.40 per million tokens and output at $4.40 per million. By comparison, GPT-5.5 sits at $5/$30 and Claude Opus 4.5-4.8 at $5/$25.
Excellent pelican, disappointing opossum
When I previously tested GLM-5.1 with the prompt “Generate an SVG of a NORTH VIRGINIA OPOSSUM ON AN E-SCOOTER”, the result was a standout piece of generative art. The model wrapped the SVG in HTML to add CSS animations, creating a dynamic scene. GLM-5.2 was tasked with a similar prompt: “Generate an SVG of a pelican riding a bicycle”. The output was a fully animated, self-contained vector illustration. The bicycle anatomy was correct, with spokes, wheels, and pedals rotating in sync, while the pelican featured a red scarf and bobbed realistically. The only flaw was that the feet did not stay on the pedals.
Unfortunately, the model struggled significantly with the opossum prompt. The resulting image featured a green scooter that lacked distinct characteristics, a possum wearing a helmet that rendered it barely recognizable, and a confusing background grid. Most notably, the model did not attempt any animation, a stark regression from the previous version.
This is a clear step backward for the opossum task. The GLM-5.1 version produced a dark, atmospheric scene where the animal was clearly identifiable, the scooter was functional, and the tail bobbed naturally. The only glitch there was the occasional eye detachment. GLM-5.2 simply failed to deliver that level of creative execution.
Key takeaways
- GLM-5.2 is currently the highest-scoring open-weights model on the Artificial Analysis Intelligence Index, but its efficiency is poor, using significantly more output tokens than competitors.
- Despite being text-only, the model ranks highly on web development benchmarks, suggesting visual input may not be strictly necessary for complex coding tasks.
- The model shows inconsistent creative capabilities; while it produced a superior bicycle illustration, it regressed sharply on the opossum prompt compared to GLM-5.1.
