Why Companies Stick with Frontier Models Despite Cheaper Alternatives

The Last Mile Bottleneck

While models like GLM 5.2 offer performance comparable to frontier models for "center of distribution" tasks—such as brochure site generation, routine synthesis, and standard coding patterns—companies struggle to switch because they are not just replacing a model call, but an entire work system. A model is effectively a brain in a jar; without a custom harness to manage memory, tool calls, and system prompts, it cannot function as a drop-in replacement for integrated services like Claude.

The Cost of Convenience

Frontier providers like Anthropic are building "sticky" team-level harnesses, such as Claude Tag, which integrate directly into Slack. These tools capture company context automatically, creating a high switching cost. By using these services, companies effectively rent their own institutional knowledge back from the model provider. Because the talent required to build custom, model-agnostic harnesses is scarce and expensive, most organizations default to the convenience of frontier providers despite the massive potential for token cost savings through open-source routing.

Strategic Opportunities for Builders

There is a significant market opportunity for agencies and developers who can build custom harnesses that decouple company context from specific model providers. The path to ROI involves:

• Auditing task distributions to identify which workloads are "center of distribution" (suitable for cheaper open-source models) versus "edge of distribution" (requiring frontier models).
• Refactoring agentic pipelines to support model-agnostic memory and tool-calling architectures.
• Implementing routing logic that dynamically assigns tasks to the most cost-effective model without sacrificing quality.
• Developing proprietary harnesses that allow teams to maintain control over their data rather than ceding it to a frontier lab's ecosystem.