GPT-4o and Claude 3.5 Sonnet compressed the frontier

GPT-4o and Claude 3.5 Sonnet made the frontier feel smaller.

The useful part was not that one model beat the other. It was that both launches pushed capable models closer to ordinary product work. Latency, price, and multimodal input stopped feeling like special cases and started feeling like route design. Once that happened, “best model” stopped being a useful phrase on its own.

I started thinking about product routes instead of model trophies.

route	what it needs	why it exists
quick support triage	speed and decent text understanding	answer fast and move on
screenshot analysis	vision plus reliable explanation	review UI states and failures
coding help	strong reasoning and patch discipline	keep diffs small and useful
long-form drafting	cheaper token economics	good enough without wasting budget

That middle layer mattered because it made capable models feel ordinary. A model did not have to be the smartest thing in the room to be the right choice. It just had to be fast enough, cheap enough, and good enough for the workflow.

The evals had to change with that. I cared less about leaderboard bragging and more about patch quality, screenshot reasoning, tool stability, and whether the route made the product calmer instead of just more impressive. A model that is slightly worse on a public benchmark can still be the better product route if it reduces latency, handles the right slice, and fails in a way the app can recover from.

That also changed how I looked at demos. A support assistant that can read a screenshot and answer in under a second is more useful than a benchmark winner that takes long enough to feel ceremonial. A code helper that writes one readable patch beats a model that produces a nicer score but makes reviewers clean up the diff. The frontier compressed because the middle became good enough to ship.