Model Analysis & Selection

There is no universally best model

“Which model is best?” has no useful universal answer. The useful question is: which model is best for the work, constraints, and risk appetite you actually have?

LLMs make bespoke, nuanced software more practical: computers can now understand plain English rather than requiring every intent to be expressed through a traditional programming language. That value only matters when the chosen model can reliably help you build what you imagine.

Choose against requirements

Decide what you want to build. The more defined the goal, the better.
Decide your privacy risk appetite and allowance.
Decide which capabilities the model needs to achieve the goal:
- intelligence;
- speed (TPS);
- modes such as text, image, and video;
- context-window size; and
- determinism.
Choose the cheapest model or models that meet every criterion and can achieve the goal reliably.

Compare, then commit

Artificial Analysis is a useful source for side-by-side model comparisons. If you genuinely do not know where to start, one of the major subscription offerings is a reasonable default:

I use Codex because OpenAI consistently delivers frontier, multimodal capability at good subscription cost efficiency and is accommodating of open-source and third-party harnesses.