OPEN-SOURCE REPRODUCIBLE BENCHMARK
The AI Coding Tax Benchmark
We ran 25 real post-cutoff compiler, SDK, and protocol breaking changes across Next.js 15, Tailwind v4, MCP Streamable HTTP, x402 EIP-3009, Vercel AI SDK v4, Prisma Edge, and Pydantic v2. Spending $0.125 total on FixLedger eliminated 476,464 retry tokens and cut debugging time from 34.1 minutes to 5.8 minutes.
Token Reduction
-90.4%
527,000 vs 50,536 tok
Net Cost ROI (Sonnet 4.5)
15x ROI
$0.125 fees saved $1.88 net
Wall-Clock Speedup
5.9x
34.1 min down to 5.8 min
Single-Session Pass Rate
100%
vs 72% with blind web search
Cumulative Token & Dollar Burn Across 25 Bugs
Reproduce locally: npm run benchmarkControl Group (Frontier Agent + Free Web Search)527,000 tokens ($2.242)
Test Group (Frontier Agent + FixLedger MCP $0.005 Unlocks)50,536 tokens ($0.363 all-in)