Charles AZAM - Blog

Code is the purest way to express complexity.

Stop reasoning blindly.

I gave Claude Fable 5 and GPT-5.6 Sol the same unpublished NP-hard optimization problem, with and without their native /goal mode.

Zhipu AI's GLM-5 leads SWE-bench and LiveCodeBench.

Deep dive into the architecture of Codex, Gemini CLI, Mistral Vibe, and OpenCode.

Claude Code (Opus 4.6), Codex (GPT-5.3-Codex xhigh), Gemini CLI (Gemini-3-Pro-Preview), and Mistral (Devstral-2) tackle a fiber network optimization problem.

The agent loop itself is ~50 lines — the real work is user isolation, session persistence, and execution backends.

We benchmarked our production RAG system across embedding models, chunk sizes, chunking strategies, and retrieval modes.

A benchmark whose test set lies in the future cannot be in any training set.