Why AI-Driven Companies Must Rapidly Incorporate New Models

In 2023, a company that had built its product on GPT-3.5 faced a choice when GPT-4 launched: upgrade and capture the performance improvement, or stay on the existing model and maintain stability. Most companies stayed put. The ones that upgraded quickly — reducing hallucination rates, improving output quality, cutting cost-per-query — pulled ahead. That gap has compounded every six months since.

We've seen more AI model launches in the past two years than in the preceding decade: GPT-4, GPT-4o, o1, o3, Claude 3, Claude 3.5, Gemini 1.5, Llama 3, Mistral Large. Each generation brought meaningful capability improvements. Each also brought significant cost reductions — foundation model pricing has dropped roughly 90% since 2023. This pace is not slowing.

The hardcoded model problem

Companies that are hardcoded to a specific model version face compounding exposure. Their costs don't benefit from pricing decreases. Their performance doesn't improve as models get better. And their engineering team is stuck maintaining an integration to a model that may be deprecated — while competitors on newer models are offering materially better products at lower unit economics.

This isn't a hypothetical risk. We've seen it play out in vendor evaluations where a company's core product relied on a model that had been superseded eighteen months prior. The customer-facing quality gap was visible. The margin compression from not benefiting from pricing improvements was quantifiable. Neither had been disclosed in the pitch materials.

Model-agnostic architecture is the structural answer

The companies positioned to benefit from model improvement — rather than be threatened by it — share a specific architectural characteristic: the model layer is abstracted. The application logic doesn't care which foundation model is running underneath it. Swapping models is a configuration change, not an engineering project.

This requires deliberate design. It means building abstraction layers between application code and model APIs, maintaining consistent prompt management and evaluation frameworks, and investing in the internal infrastructure to test new models against production benchmarks before switching. None of this is free — but the cost of not doing it is higher.

The cost curve implication

The 90% cost reduction in foundation models since 2023 has a double implication. For companies with model-agnostic architecture, falling API costs flow directly into margin improvement — the product gets cheaper to run without any change to pricing. For companies locked to older, more expensive models, competitors on newer pricing tiers can undercut them without sacrificing quality.

In markets where pricing pressure is real, this creates a structural disadvantage that compounds quarterly. It's also a disadvantage that's entirely self-inflicted — the result of architectural decisions made early that were never revisited.

What to look for in diligence

• Is this company positioned to benefit from model improvement, or threatened by it? Ask specifically: when was the last time you upgraded your primary foundation model, and how long did that take?
• Is there an abstraction layer between the application and the model API, or are model-specific calls embedded throughout the codebase? The answer to this question tells you more about the team's architectural sophistication than any amount of documentation.
• How has the company's AI cost-per-transaction changed over the past 12 months? A company on model-agnostic architecture with good engineering should be showing cost improvement even without revenue growth. Flat or rising costs are a signal worth investigating.
• Does the team have a view on the model roadmap and how their product would be affected by the next generation of capabilities? Teams that follow the model landscape and have thought through the implications are fundamentally different from teams that treat the underlying model as a fixed dependency.

The broader point

Model adaptability isn't just a technical property — it's a competitive and financial one. The companies that treat it as infrastructure, not an afterthought, are the ones that will continue to improve their product and their economics as the AI landscape evolves. The ones that don't are making a bet that the model they've built on stays competitive. That's not a bet with good odds.