GLM-5.2 from Z.ai: Chinese AI agent challenges Claude, but it's still far from being a "killer"
A new wave of competition is brewing in the world of artificial intelligence. Chinese company Z.ai has unveiled the GLM-5.2 model, which the community has already dubbed the "killer" of Anthropic's flagship Claude. Can this new neural network truly dethrone the leader, or is it just another hype?
As an analyst, I have carefully studied the technical specifications and reviews. GLM-5.2 is a flagship model aimed at solving complex, long-duration tasks. Its main advantage is a context window of 1 million tokens, five times larger than its predecessor GLM-5.1. This allows the model to keep vast amounts of code or documentation in focus without performance degradation.
Key features of the model:
- Giant context (1 million tokens): The entire codebase of a project fits into a single reasoning cycle, which is critical for complex development.
- Two reasoning modes: High — for balancing performance and token consumption, and Max — for maximum analysis depth, but with increased resource usage.
- Open MIT License: The model can be run on your own hardware, giving full control over data and costs.
- Availability: GLM-5.2 is accessible via the GLM Coding Plan subscription, the ZCode desktop agent, and is supported in Claude Code and OpenCode environments.
Benchmarks: Breakthrough or Marketing?
According to Z.ai's own tests, GLM-5.2 shows impressive results. On Terminal-Bench 2.1, it scored 81.0 points, closely approaching Opus 4.8 (85.0) and surpassing Gemini 3.1 Pro (74.0). The improvement over GLM-5.1 is colossal — from 63.5 to 81.0. On SWE-bench Pro, the model also demonstrates notable progress: 62.1 versus 58.4 for the previous version.
However, looking at other tests, the picture becomes less clear. On NL2Repo, ProgramBench, and DeepSWE, GLM-5.2 confidently outperforms GPT-5.5 and Gemini 3.1 Pro, but consistently lags behind Opus 4.8. On long-horizon tasks like FrontierSWE, the gap is minimal (only 1%), but on SWE-Marathon, the difference reaches 13%.
Benchmark conclusion: GLM-5.2 is undoubtedly the strongest open model currently available. It can compete on equal footing with closed giants in several scenarios, but it still falls short of absolute leadership, especially in the most complex tasks.
Cost and Real-World Experience
The GLM Coding Plan subscription looks attractive: starting at $12.6 per month for the Lite tier. However, as users note, "the devil is in the details." The Max mode, where the model unlocks its potential, burns tokens several times faster, making usage expensive. Criticism also targets the unstable cloud infrastructure and the model's tendency to get "stuck" in endless reasoning loops, ignoring user commands.
Many developers note that "it's easier and cheaper to pay for Claude or GPT," as Z.ai's support and service stability leave much to be desired. GLM-5.2's strengths — basic logic and persistence in achieving goals — are undermined by its weak service ecosystem.
My expert verdict: GLM-5.2 is a technological breakthrough for open models. It proves that China can create world-class AI systems. Calling it a "Claude killer" is premature. It is more of a formidable competitor that is closing the gap but has not yet bridged it. For enthusiasts and developers willing to experiment, it is an excellent tool. For businesses needing stability and predictability, Claude and GPT remain the more reliable choice. The AI market is becoming increasingly competitive, and that benefits all of us.