Founders often ask which AI model they should use, expecting a single answer. There isn't one. The best model depends on the job it needs to perform, and the right choice for a customer-facing assistant may be wrong for a background classification task in the same product.

Start with the task, not the model

Define what the model actually has to do — summarise, classify, extract, converse, write code — and how good is good enough. A clear task definition tells you far more about the right model than any leaderboard.

The dimensions that matter

Weigh a few practical dimensions against each other: capability on your specific task; latency, since a slow response can be unusable in an interactive product; context requirements, if the model must work over large documents; data sensitivity, which affects where data can be sent; provider dependency and the risk of relying on a single vendor; deployment options across hosted and open-source models; and ongoing cost at the volume you expect.

Hosted versus open-source

Hosted models from major providers offer strong capability with no infrastructure to run, at a per-use price. Open-source models can be run in your own environment, giving more control over data and cost but adding operational responsibility. Many products use a mix — a capable hosted model where quality matters most, a smaller or self-hosted model for high-volume, lower-stakes work.

Avoiding lock-in

Model quality and pricing change quickly. Designing so you can switch models — a clear boundary in your code and a way to evaluate a new model against your real tasks — protects you from betting the product on one provider's roadmap.

Cost and latency in production

A model that looks affordable in testing can become expensive at scale, and one that feels fast for a single request can degrade under real traffic. Estimate cost and latency at your expected volume before you commit, and monitor both once you are live.

Chosen deliberately, the model becomes one well-understood component of the product rather than a bet you hope pays off.