Insights

Perspectives on practical AI.

Field notes on AI adoption, implementation, and the workflows that actually ship — from the Second Wave team.

FeaturedLLM CostsClaude

We Tried to Benchmark the Viral 70% Claude Cost Trick. Claude Refused the Benchmark.

An open-source proxy claims to cut Claude API bills by up to 70% by re-rendering text as packed images. We built a harness to test it — and ran into something more interesting than the cost math: Fable 5's safety layer treating the benchmark itself as hostile.

July 7, 2026 · 8 min readRead

Latest