the tokenizer is the ISA

A coding language built for AI, not people.

AI assistants write computer code all day — and every word they write costs time and money. curt is a programming language designed so they write far less of it: the same programs, a fraction of the words. Every claim on this page is measured, and you can reproduce it yourself.

Below is a small program — a web server that echoes back whatever you send it — written in curt. The big number is roughly how many "words" (tokens) an AI spends to write it; fewer means cheaper and faster. Don't worry if the code looks alien — it's meant for an AI to write, not for you to read.

specimen · tcp uppercase echo o200k_base · real corpus program

        
hover the code or a note — they link

what this actually is

An experiment, built in the open.

I'm not a compiler engineer. curt started as a question — could a programming language be designed for the way AI writes code, so it costs less to use? — and I built it by directing AI agents to design it, write it, test it, and measure it. That's fitting: it's a language for AI, made largely by AI.

What I can stand behind is the evidence. Every number on this site comes from a script anyone can run, and the honest results are all here — including where curt loses (on small, everyday tasks, plain Python is still cheaper). Judge the results, not my résumé.

measured, never estimated

1.10× about the same token cost as Python to write — and far less than Go or Rust how ↗
0 syntax errors across 200 programs an AI wrote under curt's guard rails how ↗
7.7× cheaper per task than the closest comparable AI language (on one model) how ↗
$0 to try — the whole language runs in your browser, nothing to install play ↗

Honest by construction: on small, everyday tasks, plain Python is still the cheapest — curt's edge shows up on larger programs and in catching and fixing mistakes. The full, mixed results (wins and losses) are on the benchmarks page.

why it might matter

Built for how AI writes — and measured like an experiment.

Most programming languages are designed for people to read. curt is designed around the one thing that costs real money when an AI writes code: how many tokens it takes. Every feature had to earn its place by being measured against the alternatives — and the tidy, human-readable version is something the tool can generate for you, not something you have to write.

fewer words, by designEvery part of the language was tested for how many tokens it costs an AI to write — and the cheaper option won.
an AI can't typocurt ships with guard rails that make it impossible for an AI to produce a syntax error in the first place.
it explains its own mistakesWhen something's wrong, curt describes the fix in a form an AI can apply on its own — usually on the first try.
safe by defaultA curt program can't touch your files or the network unless you explicitly allow it.