Claude Code vs Agent Zero, same model: Claude Opus 5 High, released yesterday. Build a working Web OS from scratch. Terminal, apps, two games. Claude Code: 9:58, 80.3k tokens Agent Zero: 11:34, 59.7k tokens That's 26% fewer tokens for the same result. Full test at the end

The exact prompt used for both: "Using html, css and js, generate a browser OS with the following features: - At least 5 applications - - Two of the 5 applications must be FUNCTIONAL games - Ability to change wallpaper - A "special" feature that you decide on and document what it is & why it is special. - The result must be contained within a single script, and be able to be opened in Chrome browser. - Put all your work in a local folder called webos"

The surprising part: despite spending far fewer tokens, Agent Zero's Web OS wasn't the cheaper-looking one. Both work nearly identically. But the A0 version has a functioning file manager and even Paint. The Claude Code one doesn't (that's random and the opposite could always happen. Both scripts are one-shotted in the video). 90 extra seconds. 20.6k tokens saved. Judge the results yourself in the video.

tweet doing well ;)

that is just crazy

how about quality on serious projects ? (not single html file)

we went from "hello world" to "build me an operating system" real casually 💻

🔥🔥🔥🚀

agents speedrunning themselves out of a job 59k tokens is literally a guy who reads the recipe instead of just winging it 💀

26% fewer tokens for the same result, that's a solid efficiency gain.

26% token savings is wild for the same output.

the faster one used 35% more tokens and thats somehow the win they led with 😐

省 token 可以，慢一分半也得看场景

Best harness out there 👏

干掉大厂😄

Good stuff, I can say I have appreciated it. Keep up good work brother💪🫡🖥️

nobody asking why building a whole web OS is just a benchmark now we out here casually shipping operating systems in 10 mins like its a homework assignment 😭

Which game was harder to build? The token savings are wild, but I'm curious about the complexity distribution.

Token efficiency at this scale is huge. 26% fewer tokens on a real working OS means the agent spent less time retracing and re reading context, which usually points to tighter planning loops rather than smarter coding per se.

Agent Zero is more efficient

Same model is a good control, but “same result” still needs a rubric. Were the prompt, permissions, tool schemas, retry budget and final verification identical? Token efficiency matters only after the output passes the same acceptance tests.

Interesting comparison. Token efficiency is becoming more important than pure speed as people start running longer autonomous loops. I’ve been focusing on architecture (memory + verifier + routing) more than the model itself lately. The combination matters a lot.

Token efficiency is the metric that actually matters at scale. A 26% reduction compounds fast when you are running builds all day. Time is cheap, tokens are not.

A 26% token gap is material only if quality holds under adversarial tasks. Cheap runs that fail silently are the most expensive ones. We need outcome quality, traceability, then token efficiency.