Don't Marry a Model

OpenAI ended its exclusive deal with Microsoft yesterday. By this morning, OpenAI's models were already live on AWS Bedrock. Even the companies building AI can't stay locked in to one partner.

Meanwhile, GPT-5.5 scored 87 on a complex knowledge-work benchmark where the next closest model (Claude) scored 67. Six months ago, Claude was ahead. Now GPT-5.5 is. In another few months, something else will be.

This is the part that matters if you run a business: the leaderboard reshuffles constantly, and it will keep reshuffling. One developer recently shared how a single prompt caching miss on agentic workloads produced a $38,000 AWS bill. Wrong model, wrong configuration, no flexibility to route around it.

A practitioner analyzed over 3,000 real-world AI deployments and concluded: the harness around the model matters more than which model you pick. The architecture, the routing, the ability to swap one model for another without rebuilding your whole workflow.

The companies getting this right aren't betting on one vendor. They're routing different types of work to different models (research to one, code to another, high-volume tasks to a third) and treating model choice the way they treat any vendor contract: reviewable, replaceable, and tested against their own results.

Don't marry a model. Own the routing layer so you can move when the math changes. And it will change.