The release of the new flagship model GLM-5.2 from Zhipu AI under the MIT open license poses a serious challenge to proprietary solutions like Anthropic's Claude Opus 4.5, offering superiority in coding and significant resource savings.
What Happened
The new GLM-5.2 model, based on the MoE (Mixture of Experts) architecture with 753B parameters, has shown better results in specialized tests: in the SWE-bench Pro benchmark, it scored 62.1 points compared to 41.8 for Claude Opus 4.5, and in the Google-Proof Q&A factual accuracy test, it reached 91.2 compared to 87.0 for its competitor. Additionally, GLM-5.2 supports a context window of up to 1 million tokens.
Context
While Claude Opus 4.5 remains a leader in multimodality (Vision) and enjoys wide availability through cloud providers (AWS, GCP, Azure), the emergence of a high-performance open-source model with an MIT license radically changes the economics of LLM usage.
Why It Matters for the Industry
For the industry, this means serious pressure on the pricing of closed APIs. The use of the MoE architecture in GLM-5.2 allows for reducing token costs by up to 257% compared to Anthropic's solutions, stimulating a shift among companies toward optimized self-hosted inference stacks like vLLM or TensorRT-LLM.
Why It Matters for Users
For developers and data analysts, GLM-5.2 provides a more efficient and cheaper tool for writing code and processing ultra-long documents (up to 1 million tokens). Meanwhile, Claude Opus 4.5 is still recommended for multimodal tasks and scenarios where high fault tolerance through various cloud providers is critical.
What Is Not Yet Known / Limitations
Engineering and architectural specialists express concerns regarding the risks of using open models compared to proven proprietary solutions in corporate environments.
Sources
Author
Look at AI, Editorial Team