GPT-4o and Claude 3.5 Sonnet compressed the frontier

GPT-4o and Claude 3.5 Sonnet made frontier capability feel closer to ordinary product work.

The important comparison was not which model won. Latency, price, visual input, and coding behavior all became routing choices. “Best model” stopped being a sufficient product description.

The routes looked more useful:

route	needs	reason
support triage	speed and text understanding	answer quickly
screenshot analysis	vision and explanation	inspect UI states
coding help	reasoning and patch discipline	keep diffs usable
long-form drafting	economical tokens	avoid wasting budget

A model does not need to be globally strongest to be the right route. It needs to be fast enough, affordable enough, and reliable on the task slice that matters.

That changes evals too. Measure patch quality, screenshot reasoning, tool stability, latency, cost, and recoverability. A model that scores lower on a broad benchmark may still be the better product route if it handles the important slice and fails in a way the application can recover from.

The frontier compressed because the middle became good enough to ship. That raised the standard around the model: route by job, verify the output, and optimize the workflow rather than the trophy.