Lucena Notes

Token-efficient
AI engineering

Essays on current context, digital mise en place, and why bigger context windows are not a substitute for product architecture.

Benchmark / Token Efficiency

The Almedra Token-Efficiency Benchmark, Updated For GLM 5.2

Six coding agents, the same Almedra website task, GLM 5.2 on StreamLake, and a fresh look at tokens, cache mix, calls, and completion.

Read update ↗

Benchmark / Token Efficiency

The Almedra Token-Efficiency Benchmark

Six coding agents, one clean small-business website task, one OpenRouter model, and one third-party ledger per run.

Read benchmark ↗

Models / Token Efficiency

The Best Models for Token Efficiency When Using Lucena Coder

Our favorite models for Lucena Coder, with input pricing, output pricing, and a practical speed tier.

Read guide ↗

Ethos / AI Agents

Mise En Place For AI Agents

Everything a step needs, nothing it does not. Our clearest articulation of the Lucena Coder workbench philosophy.

Read essay ↗

Token Burn / Coding Harnesses

Where Coding Agents Burn Tokens

A tour through the hidden places coding agents waste money: stale failures, duplicate file bodies, thin receipts, broad tools, and terminal half-states.

Read essay ↗