Sources
Fireship Ran DeepSeek Harness End to End: 2.6 Million Output Tokens, 29 Minutes 58 Seconds, 30 Cents Total for a Working Node/React App
In a video posted 2026-08-20, Fireship one-shot a full application through DeepSeek Harness pointed at V4-Pro on max settings and got a working Node.js and React build for 30 cents, burning 2.6M output tokens over just under 30 minutes. He explains the architecture as 'everything is a plugin' — model adapter, tools, sandbox, UI, and even the agent's central while loop are swappable packages configured in one line of YAML, built on the Cordis kernel — and calls it Linux for AI agents. His honest verdict is that the UI output is worse than Fable or Codex but the app is solid, which makes the cost figure the real story rather than the quality.
↳ Follow the thread