KM4BI KM4BI

GLM 5.2 Is Free and Beats Claude on Most Work. Why Can’t Companies Switch?

GLM 5.2 makes cheap open-source AI look practical, but companies are held back by workflow harnesses, context, routing, and switching costs.

GLM 5.2 shows that powerful AI is moving sharply down the cost curve. Nate B Jones argues that the model is genuinely impressive on the broad middle of everyday AI work: content drafts, standard decks, routine synthesis, familiar coding patterns, brochure sites, and outputs that humans can inspect quickly.

The model is not the bottleneck

For center-of-distribution tasks, GLM 5.2 may be more than “good enough.” It can be fast, cheap, high quality, and in some cases better than Claude. That matters because most knowledge work is repetitive enough to fall into recognizable patterns.

The harder question is whether a company knows which parts of its workload fit that category. Edge cases, ambiguous tasks, and high-stakes work may still justify frontier models. Few teams currently measure their work in those terms.

Switching means rebuilding the harness

The central point is that companies are not just replacing a model call. They are replacing a work system. Prompts, memory, tool calls, routing logic, and user interfaces often depend on how Claude or OpenAI behaves. The Lindy example shows the scale of the change: moving toward DeepSeek required rebuilding the harness rather than lifting existing workflows unchanged.

That makes open-source adoption easiest when the ROI is immediate, such as in AI-native products where token savings flow directly into margins. For internal knowledge-work automation, the payoff is real but harder to quantify.

Why frontier vendors stay sticky

Claude Tag is a key example. By letting people tag Claude in Slack, Anthropic is building a team-level harness that absorbs messy organizational context. The product becomes useful precisely because it is close to where work happens.

That stickiness changes the economics. Even if GLM 5.2 is dramatically cheaper, a company may hesitate to abandon the model already embedded in its conversations, decisions, and accumulated context.

The opportunity

The next competitive layer is the last mile: model routers, team harnesses, memory systems, and agentic pipelines that can send routine work to cheap open models while reserving frontier models for harder tasks. Builders who can deliver that refactor will be valuable in 2026 and 2027, because they help companies stop renting their own context back from frontier providers.

Source