A new model has been released by Chinese developers: GLM-4.7. According to documentation, this model now emphasizes task execution rather than just code generation. It was initially reported to outperform Sonnet 4.5 and GPT-5.1 in certain benchmarks, but the documentation has since been edited to remove references to these models. The model achieved an open-source SOTA score of 84.8 on LiveCodeBench V6, surpassing Claude Sonnet 4.5, and also set a new open-source SOTA in the AIME 2025 math competition, outperforming Claude Sonnet 4.5 and GPT-5.1. It scored 42% on the Human Last Exam (HLE) benchmark, representing a 38% improvement over GLM-4.6 and approaching GPT-5.1 performance.
The user tested the model with a plan to implement backend updates with migrations and frontend development, based on GPT-5.2. Surprisingly, the model performed very well, completing the tasks in just 18 minutes using only 70,000 tokens. The previous version (4.6) often struggled and gave up early on similar tasks, requiring detailed instructions and the use of Claude code combined with Chrome DevTools to pass end-to-end tests. To the user’s surprise, the new model handled the tasks excellently, outperforming Gemini Flash 3, and completed the tests without issues, despite not supporting images. It was observed that to read content from images, the model in CC calls a specific tool that processes the image and returns the text.
The tool, analyze_image, is used as follows: Input: {“imageSource”:” “,”prompt”:”Describe what text and content is shown in this image”}. It executes on the server and provides a summary of the image content. This method works only with screenshots, as Chrome DevTools seem to send the DOM tree to the model. Overall, the first impression is very positive. The user already had a quarterly subscription to Z.ai (paying $50 for the mid-tier plan) and uses Claude code with a script that allows switching subscriptions by updating the model number and API key. The subscription is quite affordable, starting at $3 per month. There is also a referral program offering an additional $10 discount if subscribing for more than one month. The user provides links for more information:
http://z.ai/
https://docs.z.ai/guides/llm/glm-4.7
http://Z.ai/
https://maas-log-prod.cn-wlcb.ufileos.com/anthropic/c86a1d8b-….png?UCloudPublicKey=TOKEN_e15ba47a-d098-4fbd-9….&Expires=17664…&Signature=6uRZH8ppbDvh…..=
https://gist.github.com/timurkhakhalev/b36428a6b1b25c9cba446725a740e231
http://z.ai/
https://z.ai/subscribe?ic=YFISF15UJC