GLM 5.2, an open-source AI model, offers high-quality, cost-effective performance that often surpasses Claude on common tasks. However, companies struggle to switch due to the complexity of integrating new AI systems, the need for specialized 'harnesses,' and the sticky nature of existing models like Claude with team-level integrations. The future of AI adoption hinges on building these last-mile solutions and balancing cost with convenience.
I recently tried GLM 5.2, an open-source AI model, and it truly blew my mind. This article explores where GLM 5.2 can be effectively used, where it can replace expensive models safely, and why switching AI models is more complex than just replacing a single model call — it involves replacing an entire work system.
GLM 5.2 is not just cheap; it is incredibly cost-effective to run on the cloud and free if you set up your own servers. For many routine AI tasks — what I call the "fat middle" of everyday AI work — GLM 5.2 performs exceptionally well, often better than Claude.
Normal work includes tasks like:
These tasks have familiar patterns and outputs that humans can quickly verify. In this "center of distribution" work, GLM 5.2 is fast, cheap, easy to use, and delivers extremely high quality.
GLM 5.2 often outperforms Claude in these common tasks, making it arguably the best model globally for center-of-distribution tasks, especially where front-end taste matters. However, despite its strengths, GLM 5.2 is not my daily driver, and many companies hesitate to switch to it fully.
Many companies want to transition to a generic routing system that directs tasks to the cheapest model available. However, this is not easy in practice because:
Ergonomics of Work: Employees often demand access to well-known frontier models like Claude or OpenAI, which have strong brand recognition and user familiarity. This creates pressure on IT departments to maintain these models.
Task Distribution Complexity: It is difficult to determine whether a company's workload is mostly center-of-distribution (routine tasks) or edge-of-distribution (complex, unique tasks). Frontier models excel at edge tasks, while open-source models like GLM 5.2 excel at center tasks. Most companies have not yet measured their task distribution properly.
Need for Specialized Harnesses: Switching models is not just about swapping APIs. Models require their own "harnesses" — the systems that handle prompts, memory, tool calls, and context management. For example, the Lindy team, led by Flo Crivello, had to rewrite their entire system to switch from Claude to an open-source model.
Talent Scarcity: Building these harnesses requires highly skilled AI engineers, who are scarce and often command high salaries. Only large or well-funded companies can afford to build these last-mile solutions.
A model alone is like a brain in a jar — it is not useful without a harness. Harnesses are the software and infrastructure that make AI models practical and effective in real-world workflows.
Products like Claude Tag integrate deeply with team workflows, automatically ingesting context from Slack and other tools. This makes it difficult for companies to switch away, even if cheaper models like GLM 5.2 offer similar performance on many tasks.
This landscape presents a golden opportunity for agencies, consultants, and entrepreneurs who can:
The US government is currently slowing frontier model releases, creating a window for open-source models to gain traction. The key question is whether companies can develop the talent and infrastructure to build the last mile harnesses needed to leverage these cheaper models effectively.
If they can, the cost of AI intelligence could drop dramatically, reshaping the AI landscape. If not, companies may remain locked into expensive frontier models due to the convenience and stickiness of their harnesses.
GLM 5.2 is an incredible model that excels at the majority of knowledge work tasks. It deserves serious consideration by companies looking to optimize AI costs and performance. However, adopting it requires a commitment to building the last-mile harnesses that make it truly useful.
Companies and entrepreneurs should ask themselves:
This is a pivotal moment for open-source AI. GLM 5.2 has opened the door, but it is up to us to walk through it and build the future of AI work.
Good luck with that journey. Cheers!
