Crypto news

In recent days, the crypto and tech community has been stirred by news of a powerful Chinese competitor to Claude from the company Z.ai. This refers to the neural network GLM-5.2, which enthusiasts and bloggers have quickly dubbed a "killer" of Anthropic's flagship models. Let's figure out how fair these loud claims are and what the new product actually is.

GLM-5.2 is positioned as a flagship model optimized for long and complex work sessions. The key improvement over its predecessor GLM-5.1 is a stable context window of 1 million tokens, which is five times larger than the previous 200,000. This allows it to retain vast amounts of code and text in its "field of view" without losing reasoning quality, which is critical for developing complex projects.

Benchmarks: Truth vs. Hype

According to Z.ai's own tests, GLM-5.2 shows impressive results, especially on programming-related tests. The gap with GLM-5.1 is huge: 81.0 vs. 63.5 on Terminal-Bench 2.1 and 62.1 vs. 58.4 on SWE-bench Pro. On Terminal-Bench 2.1, the new model closely approaches the scores of Claude Opus 4.8 (85.0) and surpasses Gemini 3.1 Pro (74.0).

However, looking at the full picture, GLM-5.2 still lags behind Opus 4.8 on most key metrics. On the SWE-bench Pro test, the result is 62.1 vs. 69.2 for Opus, and on NL2Repo, the gap is enormous: 48.9 vs. 69.7. The exception is the Terminal-Bench 2.1 test, where the gap is minimal. On long-duration tasks (FrontierSWE), the gap from Opus 4.8 is only 1%, indicating the model's strong endurance.

The Price Question and "Pitfalls"

The main advantage of GLM-5.2 is its price. The GLM Coding Plan subscription starts at $12.6 per month (Lite tariff with annual payment), while competitors from Anthropic and OpenAI are significantly more expensive. However, as always, there are nuances.

Users are already actively testing the new model and sharing mixed reviews. The model's strengths are acknowledged: excellent basic logic, ability to autonomously perform tasks, and "persistence" in achieving goals. But criticism primarily concerns the service and stability:

• Weak cloud infrastructure: Despite a good mathematical model, the server side leaves much to be desired.
• High cost in Max mode: To unlock the full potential of GLM-5.2, the maximum reasoning mode is required, which consumes many times more tokens, negating the price advantage.
• Behavioral issues: The model is prone to infinite loops and ignoring commands, creating the impression that it is "tuned" for benchmarks rather than real code.

Conclusion: Killer or Not?

There is no clear answer. GLM-5.2 is undoubtedly the strongest open model available today for programming and autonomous tasks. It has come very close to Anthropic's flagships in certain scenarios, especially in terms of price. The open MIT license and the possibility of self-hosting make it attractive for developers who want to control their data.

However, calling it a "killer" of Claude is premature. By most tests, Z.ai itself ranks its model below Opus 4.8. Issues with cloud infrastructure and high token consumption in maximum mode make it not such a budget-friendly alternative in real-world work.

My expert opinion: GLM-5.2 is a significant step forward for open-source AI, narrowing the gap with leaders but not yet surpassing them. It is an excellent tool for developers who value control and are willing to tolerate the "rawness" of the service. But for the mass user who needs stable and predictable "out-of-the-box" performance, Claude and GPT remain a more reliable choice. The AI market is becoming increasingly competitive, and that only benefits end users.