SRSHAW RANA

M-02 · MISSION LOG · 2026-09-16

The bake-off

34/34

IN 69 S

How I cut token cost and kept Claude in charge.

8.1B
tokens, one week

One week of agents burned 8.1 billion tokens. Most of it was execution.

Spec
Second vendor
Fast model
Bigger model
Verifier34 checks

Same spec. Same 34 checks. Only the model changed.

Fast model69 s · 34/34
Bigger model79 s · 34/34
Second vendor189 s · 34/34

All three passed. The clock decided it.

2.7x
faster
$0.03
per run

Same score. Faster, cheaper, and no usage cap.

0 s0/34

Pick a lane. Watch the verifier run.

Claudeplans
Fast modeltypes
Claudereviews

A cheap model does the typing. Claude still plans and reviews every result.

This lane built most of the site you're on.

← MISSIONS

M-02 · 01 / 08