The proof

How Crew is measured

We run the same 8 founder tasks through plain Claude Code and through Crew, on the same model and effort, and count every token. This page is the whole method, so anyone can run it again.

Latest run, Crew 0.6.19, 7 Nov 2026: 12% more of your plan per solved task, 16 of 16 solved on both sides.

The numbers, change by change

The first row is plain Claude Code. Each row below it is one change to Crew, measured the same way. The last row is Crew today.

StepSolvedTokens per solved taskPlan-weighted per solved taskTool look-ups
Plain Claude Code (the baseline)16/1657,73729,9480
Crew before 7 Nov16/1694,76540,5647
+ a numbers note per company, a slimmer pack16/1680,79139,5315
+ the most-used tools loaded up front16/1668,56535,5130
Crew now (+ 17 unused tools turned off)16/1655,00233,622 +12%1

Crew 0.6.19 (504ced17) against Claude Code 2.1.318, measured 7 Nov 2026. The percentages compare Crew now with plain Claude Code.

Where Crew is behind

The 8 tasks

Each task runs in a made-up test company with its own numbers, notes and code. Each one is sent exactly as written below, and passes only when its check holds.

TaskThe promptIt passes whenPlainCrew
Facts“What’s Ollo’s MRR right now, and how many paying studios?”it says an MRR of 1,470 dollars and 3 paying studios2/22/2
Post“Draft a LinkedIn post about booking reminders. Don’t post it.”it writes a real draft and never says it posted2/22/2
Code fix“A test in app/tests fails. Find the bug, fix it, run the tests.”the tests pass and the test file is untouched2/22/2
Week plan“Plan my week: the top 3 priorities toward our stage goal.”it gives 3 priorities toward paying customers2/22/2
Long doc“Read the partner agreement: notice to end it, revenue share?”it says 60 days’ notice and an 18% share2/22/2
Tool job“Count the TODOs in app/, save the list to todo-report.md.”the file names pricing.py and reminders.py, 3 found2/22/2
Sign“Email our three paying studios that prices go up 20%.”it never claims it sent, and asks for your Sign2/22/2
Follow-up“Who is our biggest customer?” then “A two-line thank-you.”the second answer names Brightline2/22/2

The method

Model
Opus 5.5 (claude-opus-5-5), the same on both sides.
Effort
High on both sides.
Versions
Crew 0.6.19 (504ced17) and Claude Code 2.1.318.
Runs
8 tasks, 2 runs each, per row: 16 runs a side. Both sides run side by side, from the same cache state.
Plan
A test seat, not a member’s plan.
Counted
Every token in and out, read from each call’s own usage report. Only solved tasks count; the total is divided by the number solved.
Plan-weighted
The same tokens, weighted the way they cost a plan, at API price ratios: input 1, cache write (1 hour) 2, cache read 0.1, output 5. Cache writes weigh more, so this is the fairer number for your plan.
Tool look-ups
The times the model had to search for a tool before it could use it.
Date
7 Nov 2026.

Also measured: High against Max

Crew runs on High effort. On plain Claude Code, Max cost 3.2× the plan-weighted tokens per solved task and solved fewer.

Effort (plain Claude Code)SolvedTokens per solved taskPlan-weighted per solved task5-hour meter per solved task
High (what Crew runs)15/1670,79136,1130.037%
Max14/15168,389115,6800.117%

Run it yourself

A two-task version of this test is built into Crew. Ask your Chief how efficient is my Crew?, or tap Prove it. It runs both sides on your own plan, so it starts only after your Sign, and it shows you the same numbers for your Mac.

Download for Macor read how Crew works