
I Analyzed 165.9 Million Tokens to Understand How Claude Code Actually Works
A deep dive into Claude Code's internal architecture - real JSONL session data, 165.9M tokens, multi-model routing, and why 97% of all tokens are cache reads.
Deep dives, system design, and lessons from building commerce infrastructure at scale. By the engineers at devx.

A deep dive into Claude Code's internal architecture - real JSONL session data, 165.9M tokens, multi-model routing, and why 97% of all tokens are cache reads.

I built a 3D diamond ring configurator — 139,968 variants, physically accurate ray bouncing, real chromatic dispersion — using Claude Code, Manus, and ChatGPT as my engineering team.


How a 150-person engineering shop went from AI tool fragmentation to a shared mental model for Claude Code: skills, subagents, and what's next.


Stop looking for the 'best' AI coding assistant. Strategic tool switching saved us 5 weeks on a full-stack 3D platform.
Introducing the devx Engineering Blog - a space where we document the real problems we solve, the systems we build, and the lessons we learn along the way.
