ScaffBench

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

Read the blogRun it yourselfTry out MCP

Validation passing

200k150k100k50k0k0%25%50%75%100%cheap + reliable ↗Avg output tokens per scaffoldGPT-5.5 · high54%26.0kFable 5 · high50%61.6kFable 5 · low46%22.0kOpus 4.8 · max27%148.0kOpus 4.8 · low27%21.5kSonnet 5 · max20%113.3kSonnet 4.6 · high9%38.0kCodex Spark · high9%59.9k

Not metered on this axis (excluded from the plot): Gemini 3.5 Flash · high

Pass 1 by model

Prompt · Core + Full pass, wired libs & time

ModelCoreFullWiredTimeAvg costOut tokSteps
GPT-5.5[high]
54%54%93%6.8m$1.4126.0k14
Fable 5[high]
50%42%97%10.6m$5.1461.6k50
Fable 5[low]
46%31%95%4.0m$2.1222.0k37
Opus 4.8[low]
27%18%96%5.5m$1.6321.5k38
Opus 4.8[max]
27%27%97%36.4m$9.21148.0k99
Gemini 3.5 Flash[high]
27%27%94%111s——0
Sonnet 5[max]
20%20%96%35.5m$5.26113.3k215
Sonnet 4.6[high]
9%9%94%11.4m$1.5338.0k46
Codex Spark[high]
9%9%92%120s—59.9k12
DeepSeek V4 Flash[default]
10%10%92%4.3m—22.8k45
MiMo V2.5[default]
0%0%91%4.1m—19.4k44
Nemotron 3 Nano 30B[default]
0%0%80%3.1m—13.7k16
Nemotron 3 Ultra 550B[default]
0%0%34%8.9m—8.4k18
0%20%40%60%80%100%

Give your agent the fast path.

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

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

run in your terminal

Patreon