Frontier model

The term is used by labs, researchers, and policymakers to denote models at or near the capability ceiling — the flagship releases from a handful of labs training at the largest scale. Frontier status is inherently temporary: each generation is overtaken within months, and yesterday's frontier capability reappears in mid-tier models at a fraction of the price.

Frontier models carry the highest per-token rates in the market and often add expensive reasoning modes on top. That premium is justified for genuinely hard tasks — complex coding, deep analysis, high-stakes agents — and wasted on routine ones.

A common cost pattern is frontier-last routing: default to a cheaper tier, escalate to the frontier model only when the task demands it or a first attempt fails, and re-benchmark regularly because the price of last year's frontier performance falls rapidly.