Why Can't Companies Switch to Cheaper AI Models?

TL;DR

Companies cannot switch models easily because they are replacing an entire work system, including context, prompts, memory, tool calls, routing, and user workflows. GLM 5.2 can outperform Claude on many routine, easily checked tasks at far lower cost, but frontier models and their integrated harnesses remain valuable for unusual tasks and deeply embedded team workflows.

Transcript

I tried GLM 5.2 and it blew my mind. By the end of this video, you should know where GLM 5.2, an open-source model, can be clawed, where it can safely replace an expensive model, and where switching models is a bit of a trap because you're not replacing a model call. You're actually replacing a whole work system. And that's the thing I want to draw... Read More

Key Insights

• GLM 5.2 is presented as exceptionally strong for center-of-distribution work, including brochure sites, standard presentation outlines, first-pass copy, routine synthesis, and familiar coding tasks. These jobs follow common patterns, have many precedents, and produce results that people can evaluate quickly.
• Center-of-distribution tasks are common, recognizable problems whose expected answers and formats are relatively normal. The transcript argues that open-source models can excel in this area because similar requests have been attempted many times and their outputs are comparatively straightforward to inspect.
• Edge-of-distribution work is a stronger reason to retain frontier models because unusual tasks do not fit established patterns as reliably. Companies therefore need to measure the actual distribution of their workloads before deciding which requests can safely move to a cheaper model.
• Switching models is a systems migration rather than a simple change to an API call. Prompts, memory architecture, tool calling, context management, routing, and surrounding workflows may all require redesign because a harness optimized for one model may not transfer effectively to another.
• The business case for rebuilding a harness is clearest when AI token costs directly affect a product's margins. Lindy reportedly rebuilt its architecture around DeepSeek because cheaper, effective AI service could return savings directly to the company, while internal automation teams may have weaker incentives.
• A model without a suitable harness is described as an impressive brain in a jar. Practical value depends on the system that supplies context, manages memory, invokes tools, coordinates work, and presents the model through an interface that employees can use without substantial friction.
• Team-level harnesses are becoming strategically important because organizations already have individually productive AI workflows but still need to coordinate them. Claude Tag makes Claude available through Slack, allowing ordinary knowledge workers to request work inside a shared environment that already contains organizational context.
• Organizational context can become a source of provider lock-in even when a competing model is about 98% cheaper. If an incumbent assistant has already absorbed the usable context of Slack workflows, replacing it means recreating access, convenience, and continuity, not merely selecting another model.

Questions & Answers

Q: What kinds of work can GLM 5.2 handle well?

GLM 5.2 is presented as especially effective for routine work with familiar patterns and easily checked outputs. Examples include creating brochure websites, drafting standard PowerPoint outlines, producing first-pass copy, synthesizing ordinary material, and solving coding tasks based on familiar problem types. The transcript characterizes this as center-of-distribution work, where many examples exist and humans can quickly judge whether the output is useful.

Q: Why can GLM 5.2 outperform Claude on routine tasks?

The transcript argues that GLM 5.2 is particularly strong on common tasks with recognizable answer patterns, numerous prior examples, and outputs that are easy to inspect. It is described as producing extremely high-quality results for this middle portion of everyday AI work, especially when front-end taste matters. The claim is not merely that the model is adequate, but that it can be better than Claude for many such tasks.

Q: Why do companies struggle to replace frontier AI models?

Companies struggle because replacing a model often requires replacing the work system surrounding it. Existing prompts, memory systems, tool calls, routing logic, interfaces, and context-management practices may be tailored to the incumbent model. Employees may also prefer familiar products such as Claude or OpenAI tools. A cheaper model call therefore does not automatically reproduce the convenience, accumulated context, or reliability of the existing workflow.

Q: What is the difference between center and edge-of-distribution AI work?

Center-of-distribution work consists of common tasks with familiar structures, many examples, conventional answer patterns, and outputs that people can verify quickly. Edge-of-distribution work involves less familiar or more unusual problems for which established patterns are less dependable. The transcript recommends open-source models for much of the first category, while arguing that frontier models remain preferable when a workload is weighted toward the second category.

Q: Why must companies measure their distribution of AI tasks?

Companies must understand their task distribution because model selection should reflect the work being performed. If most requests involve common, easily inspected patterns, an open-source model may provide strong quality at much lower cost. If many requests are unusual edge cases, frontier models may remain necessary. The transcript says individuals and teams are generally not accustomed to classifying work this way, making sound routing decisions difficult.

Q: Why does changing AI models require a new harness?

Different models may require different prompts, memory designs, tool-calling behavior, and context-handling systems. The Lindy example illustrates that an organization could not simply transfer its Claude-oriented setup to DeepSeek and expect equivalent performance. It reportedly had to rebuild its harness substantially. This means migration costs include engineering and workflow redesign, not just connecting a cheaper model to the same application interface.

Q: How does Claude Tag create organizational lock-in?

Claude Tag lets people invoke Claude within Slack, creating a team-level harness where ordinary knowledge workers can request assistance without leaving their shared communication environment. Claude can receive the messy organizational context already present there, reducing the need for users to restate background information. That convenience and accumulated context make the product sticky because a replacement must reproduce the workflow and contextual access, not only match answer quality.

Q: When is switching to an open-source model worth the effort?

Switching is most compelling when an organization has a large volume of common, verifiable tasks and token costs directly affect the economics of a customer-facing AI product. In that situation, rebuilding prompts, memory, tools, and routing can produce savings that return directly to the business. The incentive is weaker for internal coding or back-office automation when the financial benefit does not clearly outweigh migration and maintenance work.

Summary & Key Takeaways

• GLM 5.2 performs exceptionally well on common, familiar tasks such as brochure websites, standard presentation outlines, first-pass copy, routine synthesis, and coding problems with recognizable patterns. These center-of-distribution tasks have abundant examples and outputs that humans can inspect quickly, making the open-source model a credible replacement for expensive frontier models in suitable workflows.

• The difficulty of switching is not limited to model quality. Organizations must determine whether their workload consists mainly of common tasks or unusual edge cases, then rebuild prompts, memory systems, tool calls, routing, and other infrastructure for the replacement model. That investment is easiest to justify when lower token costs directly improve product margins.

• Team-level harnesses create significant lock-in by embedding AI within daily workflows and accumulating organizational context. Claude Tag, for example, lets Slack users invoke Claude directly while providing access to messy conversational context. A cheaper model may answer similarly, but replacing the surrounding convenience and context requires companies to reconstruct an entire working environment.