As models churn weekly, the router becomes the durable asset
With four credible model families shipping inside a month, teams that hard-code a vendor are committing to a leaderboard position that will not survive the quarter. The infrastructure that decides which model handles which request is turning into the part worth owning.
Routing looks like plumbing, which is why it is undervalued. It is also where the durable decisions live: cost policy, privacy boundaries, fallback behaviour, and the accumulated evaluation data showing which model actually performs on your workloads rather than on a public benchmark.
That evaluation history is the real moat. A team that has measured six models against its own tasks for a year knows something no leaderboard can tell it, and that knowledge survives every model release. The models are rented; the measurements are owned.
It also explains why harness design increasingly outweighs model choice in practical rankings. A well-built harness with per-subtask model control can outperform a better model driven badly — and unlike the model, the harness is something a team can keep improving without waiting for anyone else to ship.
LLM Stats — AI updates today (August 2026) → · AI Release Tracker — Complete LLM timeline 2022-2026 →