For the past few weeks, I've been simultaneously testing GLM4.6 and Claude 4.5 on OpenRouter. My use cases are quite specific: critically evaluating published AI literature, researching my own architectural ideas, summarizing long articles, and extracting key points from lengthy conversations.

What impressed me most about GLM4.6 is its strict adherence to instructions. It can accurately understand my nuanced requirements for data analysis without adding much of its own subjective interpretation. It performs particularly well in "role thinking"—I frequently use role prompts to process information in parallel from different perspectives, such as one run dedicated to critiquing literature quality and another run seeking creative inspiration. This model responds very precisely to system prompts.

However, GLM4.6's tonal style tends to drift during role-playing, which might be a common characteristic of MoE (Mixture of Experts) models when processing role prompts. But in terms of memory and tracking of technical details, it performs more clearly than Claude, especially when processing implementation ideas or reading implementation details.

Claude 4.5's strength lies in its ability to maintain focus on complex topics over extended periods. Among all the LLMs I've tested, it has the best long-context coherence. However, its instruction-following ability diminishes over time, and it can't shake off that "professor dad" lecturing tone—much like how Gemini always comes across as an Ivy League graduate with imposter syndrome.

GLM4.6's coherence is slightly inferior, occasionally experiencing issues like Chinese language intrusion, confusion between reasoning and response layers, and repetition in long outputs. Nevertheless, it maintains consistency better than Gemini 2.5 Pro.

Most surprisingly, GLM4.6 has barely perceptible ideological bias. Compared to DeepSeek and Kimi K2, which show clear tendencies, GLM4.6 might currently be the most flexible neutral model.

If the issues with Chinese language intrusion and repetitive loops can be resolved, GLM4.6 would undoubtedly become my first choice for work. Of course, I'm still looking forward to 50B+ parameter dense models like Gemma 3 or Gemma 4.


A noteworthy detail during testing: when using through OpenRouter, it's important to confirm the quantized version of the inference service. Some users have reported subtle differences between GLM-4.5-Air-FP8 and BF16 versions in long-context scenarios, so ensuring you're using the BF16 version of GLM-4.6 is crucial for optimal performance.

Additionally, Claude's data policy changes are also worth noting. Anthropic began training on user data as of September 28th, unless manually disabled in settings. While this option existed before, they're now seriously utilizing this data. In contrast, Gemini Workspace explicitly states it doesn't train on user data, which is more privacy-friendly for users with high privacy requirements.

Overall, if you need a stable, neutral AI assistant that excels at technical analysis, GLM4.6 is worth trying. It might not have Claude's longer attention span, but its reliability in professional scenarios is impressive.