ScaffBench

Measuring coding agents on real fullstack scaffolding tasks — time, tokens, cost, and whether the result actually builds.

TokensOutput

Avg output tokens generated per scaffold run.

CostUSD

Avg API cost per scaffold in USD.

StepsActions

Avg tool steps and commands executed per run.

CodeLoC

Avg lines of code written per scaffold, lockfiles excluded.

ScaffBench 3

Full pass rate

Leaderboard

Scored across 13 specs on a clean machine. Higher pass rate is better; lower cost, time, and tokens are better.

PassPrimary

Installs, builds, type-checks, and clears all quality gates cold.

WiredLibs

Required spec libraries present and imported.

StatsMean / run

Mean time, cost, tokens, steps, and LoC.

ModelPassCoreWiredTimeAvg costOut tokStepsLoC

1GPT-5.6 Sol[high]legacy

44%67%96%10.0m$2.1528.7k215.2k

2Fable 5[high]legacy

31%31%95%10.2m$4.9458.8k48–

3Opus 5[high]legacy

31%77%96%23.8m$8.34108.8k1089.5k

4Fable 5[low]legacy

23%38%95%4.0m$2.1222.0k37–

5Ox Alpha Free[high]

15%62%98%29.2m$0.0033.8k–31.2k

0%20%40%60%80%100%

One MCP server, every spec-to-scaffold tool the benchmark used. Pick your agent, paste, done.

$ claude mcp add --transport stdio better-fullstack -- npx -y create-better-fullstack@latest mcp

run in your terminal