Back to Insights
·4 min read·Adam Roozen

The model you choose today won't be the model you run in 18 months

In 2023, OpenAI had 50% of enterprise LLM spend. By April 2026 that number is 27%. Anthropic is at 40%. The lesson isn't about which model to pick.

In 2023, OpenAI held 50% of enterprise LLM API spend. As of April 2026, that number is 27%. Anthropic, which barely existed as an enterprise vendor three years ago, now holds 40%. This shift happened without any dramatic model failure, without a major scandal. Nobody declared a winner.

The lesson isn't that Anthropic is better than OpenAI. It's that enterprise AI market share can move 20 points in two years. And it will keep moving.

The question isn't which model to build on. It's whether your architecture survives model churn. Most don't.

Why this matters for your build decisions

Every enterprise AI deployment right now makes a build-on-which-model decision. Most teams treat it as a long-term choice. GPT-5.5 launched April 23. Claude Opus 4.7 now offers 1M context with no long-context surcharge. Gemini 3.1 Pro leads on ARC-AGI-2 benchmarks. These aren't stable rankings. The frontier moves every few months, and pricing and performance follow.

The lock-in cost

When you build prompt logic, memory architecture, and evaluation frameworks tightly around a specific model, you're making a structural bet that your vendor today will be the right vendor in 2028. That bet is almost never right. The best-in-class model right now won't be best-in-class in two years. It wasn't in 2024. It wasn't in 2023.

What OpenAI's AWS move tells you

In April 2026, OpenAI ended its cloud exclusivity with Microsoft and extended to AWS. They did it because enterprise customers required it. Even the model providers are treating portability as a requirement. That's a signal worth paying attention to.

What model-agnostic architecture actually looks like

You abstract the model layer from the application layer. Prompts are versioned and tested against multiple models. Evaluation is model-independent. Switching providers doesn't require a rebuild. It requires a config change and a validation run. This isn't harder to build. It just requires deciding to build it this way from the start.

The practical test

Ask your engineering team: if we needed to swap our primary model next quarter, how long would that take? If the answer is months, or we'd need to rebuild, you've built in a lock-in problem. If the answer is a sprint, you're in a reasonable position.

Model choice matters. But model portability matters more. The enterprises that'll come through the next few years cleanly are the ones treating models as swappable infrastructure, not permanent commitments.

Written by

Adam Roozen

Strategic Advisor. AI Strategy, Digital Commerce, Technology Transformation

Nearly 30 years of operating experience · Walmart · Sam's Club · Echidna

Work with Adam