August 10, 2026
Token Wars: nerds in meltdown
What's the best programming language for coding agents?
Coders fight over the ‘best’ AI language — and the comments get way messier
TLDR: The article argues that viral claims about some programming languages being far cheaper for AI are probably based on weak tests and broken scoring. In the comments, readers split between “test real-world tasks instead” and “the savings are too small to matter anyway,” with bonus snark about the blog’s ugly design.
A fresh debate over the “best” programming language for AI coding assistants has turned into a full-on comment-section cage match. The original article pushes back on a popular idea: that shorter, more compact languages automatically save AI money and work better because they use fewer text “tokens” — the chunks of text an AI reads. Earlier writeups had claimed some quirky, compact languages crushed heavyweights like Rust or C by huge margins. But this author basically said: hold on, those tests were tiny, sloppy, and maybe measuring the wrong thing entirely.
That’s where the community pounced. One camp said the whole conversation is missing the obvious: compare languages on real projects people actually build, not toy problems that fit in a tweet. Another group argued the token gap now looks so small that it’s hard to justify skipping safer, more reliable languages just to save a few AI text chunks. In other words: the dream of a magical “cheap language” may be getting roasted in public.
And then, because the internet refuses to stay on topic, one of the biggest laughs came from someone praising the writer while dragging his blog design: great ideas, painful reading experience. Even the skeptics got spicy, poking holes in the claim that less typing always means lower AI cost, because symbols and weird syntax may actually confuse the AI’s text-splitting. The vibe? Less ‘we found the winner,’ more ‘your benchmark is lying to you.’
Key Points
- •The article reviews published claims that dynamic or concise programming languages are more token-efficient than static languages for LLM coding tasks.
- •It cites one comparison reporting a 2.6x token-efficiency gap between C and Clojure, and a later result showing J at 70 average tokens versus Clojure’s 109.
- •The article argues that the first experiment relies on trivial Rosetta Code tasks, making its results poor evidence for general conclusions about coding-agent performance.
- •It says a second benchmark was compromised by a broken test path and a symlink workaround that caused later tests to run the wrong executable.
- •The author proposes running new evaluations and pre-registers expectations, including a high-confidence prediction that the broad dynamic-versus-static claim will not hold.