The counter-trend to "bigger is better"
The module "The transformer: the paper that changed everything" showed how scaling laws drove the language-model race. In parallel, a second research branch is growing: how can you get the same performance out of a smaller, cheaper model? Techniques like "distillation" (a small model learns from a large one) and quantization (less numerical precision at almost no performance loss) make that possible.
Why this is actually the more relevant curve
For most business applications, what matters isn't absolute peak performance, but the best ratio of cost, speed, and sufficient quality (see "Token economics: how AI costs actually arise"). A small, specialized model that reliably solves a narrow task often beats a large general-purpose model - at a fraction of the cost.
Local and specialized models
Smaller models are increasingly able to run locally too - on your own hardware instead of the cloud (see "Cloud vs. on-premise: where does your AI model run?"), which strengthens data sovereignty and independence. At the same time, specialized models for narrow task areas are emerging instead of one general-purpose model for everything.
The limit: not every task can be shrunk
For tasks that need broad world knowledge or complex multi-step reasoning, a large model remains superior - shrinking works mainly for narrowly-scoped, recurring tasks. Realistically, the cost curve for standard tasks keeps shifting downward, with no fixed point where it "stabilizes" - the trend itself is more reliable than any single price quote.
Why this matters for you as a decision-maker
Anyone planning an AI solution today should factor in the cost curve, not just the current price: what's only possible today with the most expensive model can run on a small, cheap model in a year or two. That keeps reshaping build-vs.-buy (see "Build vs. buy vs. API: what's the right call?") and cloud-vs.-on-premise calculations.