I’ve been tracking model pricing for the past eighteen months. Here’s the chart that keeps me up at night, but not for the reason you’d think.
Model capability: improving ~2x per year. Model pricing: dropping ~10x per year. What cost $100 in API calls in early 2025 costs $10 today and will cost $1 next year.
This should terrify every company whose product is a wrapper around an LLM API. And it should excite every company that’s building something the model alone can’t provide.
The wrapper problem
If your AI coding tool’s value proposition is “we give you access to Claude/GPT with a nice UI,” what happens when every competitor has the same model at the same price with a comparable UI?
This isn’t theoretical. It’s happening right now. Cursor, Copilot, Claude Code, Windsurf; they all use the same models from the same providers. When Anthropic ships a better Claude, every tool using Claude gets the same upgrade. When OpenAI drops prices, every tool using OpenAI gets the same margin improvement.
The rising tide lifts all boats. Nobody gains a relative advantage from a model upgrade.
I keep watching companies in this space differentiate on UX, integrations, and features. Better editor integration. More agent types. Smoother onboarding. These matter, but they’re not moats. They’re table stakes that get copied in months. I’ve watched features ship on one platform and appear on three competitors within a quarter.
What can’t be copied
I’ve been thinking about this for a while, and I keep landing on the same answer.
Your model is a commodity. Your UX can be cloned. Your integrations are reversible. What can’t be copied: what your AI knows about your organization.
The decisions your team has made over the past year. The error patterns specific to your codebase. The conventions that emerged from how your engineers actually work. Who knows what about which systems. The record of every task the AI completed and what it learned.
This knowledge takes time to build. It compounds with every session, every engineer, every task. It can’t be exported to a competitor. And it becomes more valuable the longer you use the system.
That’s a moat. Not a feature advantage, a structural one. Not something you build in a quarter, something that accumulates over years.
The multi-model test
Here’s my test for whether an AI tool depends on the model or on its own value: can you swap the model without losing value?
If the product breaks when you switch from Claude to GPT, or degrades when you use a cheaper model for simple tasks, the value is in the model. You’re renting differentiation from your provider.
If the product works with any capable model, routing expensive tasks to powerful models, simple tasks to cheap ones, maintaining the same organizational knowledge regardless of which engine processes the request, the value is in the knowledge layer.
I think the right architecture treats models as commodity compute. Use the best model for each task. Let the knowledge layer make every model smarter about your specific organization. When a new model ships, from any provider, your product gets better immediately, without losing anything.
We’ve been running five different model providers through the same knowledge layer. The models are interchangeable. The knowledge is not.
The compounding math
Model capability improves in steps. A new release comes out, everyone gets it, the advantage resets.
Knowledge compounds continuously.
Month 1: 50 knowledge items. The AI knows some decisions and patterns.
Month 6: 300 items. It knows your architecture, conventions, and error patterns.
Month 12: 600+ items. It knows your organization’s history, expertise distribution, operational patterns.
A competitor entering at month 12 starts at zero. Same model capability. Same pricing. Their AI knows nothing about your organization. The ramp cost isn’t their product, it’s rebuilding 12 months of accumulated intelligence.
And the switching cost isn’t artificial. It’s not lock-in through proprietary formats or data hoarding. The customer built this knowledge through their own work. It’s genuinely valuable. Leaving means leaving behind everything the system learned about your organization.
Where I think the market goes
The AI coding tool market will stratify into two tiers:
Commodity tier: Model wrappers with nice UX. They compete on price, integrations, and model access. Margins compress as model costs drop. Differentiation erodes quarterly.
Intelligence tier: Systems with organizational knowledge layers. They compete on how much they know about your organization. Switching costs are organic. The product gets more valuable over time, not less.
The commodity tier is a race to the bottom. The intelligence tier is a compounding advantage.
The question for any team evaluating AI tools isn’t “which model does it use?” That’s increasingly irrelevant. The question is: what will this tool know about my organization in 12 months that it doesn’t know today?
If the answer is “nothing,” you’re buying a commodity.
About the author: Avi Cavale is the founder of Quarterback, the AI development platform that learns how your team builds. He is a serial entrepreneur, and a visionary and goal-oriented technology leader with demonstrated experience in planning, development and implementation of cutting-edge information solutions to address business opportunities. Cavale can be found online at LinkedIn.
