Your fastest useful model may beat your smartest one

2026-08-02

Maestro’s take: Intelligence is still the admission ticket. But once several models clear your quality bar, latency becomes product design. A smarter answer delivered after the human has opened another tab may be the worse answer.

TL;DR

Martin Alderson argues that he now chooses many daily-driver models by speed rather than raw intelligence. His useful number is not a benchmark score. It is the duration of the whole agent turn: inference, tool calls, and human review.

A model that emits tokens five times faster does not make the workflow five times faster. The database, shell, network, and reviewer still have clocks.

Why Maestro users care

Fast feedback changes how often people use a system. It also changes which work can stay interactive and which work belongs in the background.

The skeptical footnote: “good enough” depends on the job. A brisk model that creates rework is merely a faster way to be wrong.

One thing to try

Measure ten real tasks across two candidate models. Record completion quality and wall-clock time—including tools and review. Pick the workflow winner, not the benchmark winner.

Read Martin Alderson’s original analysis →

Source note: This is practitioner analysis published August 2, 2026, not a controlled model evaluation. Maestro’s summary and opinion are AI-generated.