Crypto news

A new loud contender for leadership has emerged in the artificial intelligence market. The Chinese neural network GLM-5.2 from Z.ai has sparked heated discussions, with many users already rushing to call it a "killer" of Anthropic's flagship model, Claude. But let's figure out how true this claim is, based on hard numbers and real reviews.

What does GLM-5.2 offer?

The main innovation of GLM-5.2 is a giant context window of 1 million tokens that does not degrade during long sessions. This is five times larger than its predecessor, GLM-5.1. Essentially, the model can keep the entire codebase of a large project in view, which is critically important for autonomous development. The model offers two reasoning modes: High — for a balance of performance and cost, and Max — for maximum depth of analysis, but with increased token consumption.

Key features that make GLM-5.2 attractive:

• Open MIT License: Anyone can deploy the model on their own hardware (self-hosting), which removes concerns about confidentiality and dependence on a cloud provider.
• Price: The API cost remains at the level of the previous version, making it extremely competitive compared to Western counterparts.
• Availability: The model is available through the GLM Coding Plan subscription (from $12.6 per month with annual payment) and is integrated with popular environments like Claude Code and OpenCode.

Numbers don't lie: what do the benchmarks show?

According to Z.ai's own tests, GLM-5.2 is the strongest open model on the market. However, it still falls short of the closed giant Claude Opus 4.8. The gap is noticeable, but it is narrowing.

Here are the results on key tests (Max mode):

• Terminal-Bench 2.1: GLM-5.2 (81.0) came very close to Opus 4.8 (85.0), surpassing Gemini 3.1 Pro (74.0) and GPT-5.5 (84.0).
• SWE-bench Pro: Here the gap with Opus 4.8 (69.2) is more significant — GLM-5.2 has a result of 62.1, which, however, is significantly better than GPT-5.5 (58.6).
• DeepSWE: GLM-5.2 showed an impressive jump from 18.0 (in 5.1) to 46.2, but still lags behind Opus 4.8 (58.0) and GPT-5.5 (70.0).

On long-horizon tasks, the picture is similar. On the FrontierSWE test, where the model works for tens of hours, GLM-5.2 lags behind Opus 4.8 by only 1%, but on the ultra-long SWE-Marathon, the lag reaches 13%. Nevertheless, it consistently outperforms all other open models.

Real reviews: euphoria and disappointment

Users are divided into two camps. Some are delighted with the logic and persistence of the model in achieving its goal, comparing it to GPT-5.5 at a high level of complexity. Others harshly criticize the service.

Strengths according to reviews:

• Best open model for programming and autonomous tasks.
• Noticeably improved basic logic compared to GLM-5.1.
• Ability to independently identify and fix errors in code by calling auxiliary agents.

Main complaints:

• Unstable cloud infrastructure: Users complain about slow performance and weak support.
• High cost: In Max mode, token consumption skyrockets, making usage expensive. Many note that it is easier and cheaper to pay for Claude or GPT.
• Behavioral issues: The model tends to get stuck in infinite loops and ignore user commands. There is an opinion that it is "tuned" exclusively for benchmarks, not for real work.

Analyst's verdict

GLM-5.2 is undoubtedly a breakthrough for open models and a serious step for China in the AI market. It demonstrates impressive results on tests and offers a unique combination of a huge context, an open license, and a low price. However, calling it a "killer" of Claude is premature. By most metrics, it still lags behind Anthropic's flagship, and the real user experience is marred by infrastructure issues and high resource consumption.

My professional opinion: GLM-5.2 is a powerful tool for developers who value open-source and are willing to tolerate the current shortcomings of the service. It creates healthy competition and pushes the leaders to evolve. But for the mass user who needs stability and predictability, Claude or GPT still remain the more preferable choice. The real battle for AI leadership is yet to come.