AI Coding Assistants Showdown: 13x Token Consumption Gap Between Codex and DSH

The author tests five AI coding harnesses paired with their respective models on a pure code-generation task: creating a 20-second teardown video explaining why claw machines rarely grab prizes. The prompt uses a /teardown skill that turns any machine into a 3D simulator, installed via npx skills add elvarofficial/teardown. Each harness gets one attempt.

Results Summary

Codex + GPT-6 Astra : 8 minutes, 3.5M total tokens (97% cached), 28K output tokens.

Claude Code + Opus 5.5 Medium : 17 minutes, 12.15M total tokens (96% cached), 130K output tokens.

Gemini + Gemini 3.8 Flash : 21 minutes, 13.94M total tokens (96% cached), 81K output tokens.

Kimi Code + Kimi K3.1 Pro : 36 minutes, 9.5M total tokens (98% cached), 59K output tokens.

DSH + DeepSeek V4.1 Flash : 46 minutes, 46.44M total tokens (99% cached), 281K output tokens.

A bar chart illustrates the efficiency gap: Codex is the most efficient, DSH the slowest and most token-hungry.

Detailed Observations

1. DSH + DeepSeek V4.1 Flash

DSH in standard mode took 46 minutes. The author notes it "overthinks" — a consistent pattern from prior tests — and even its minimal mode "thinks too much." The output video had several flaws and was the only one that responded in English despite a Chinese prompt.

2. Kimi Code + Kimi K3.1 Pro

Kimi K3.1 Pro, previously tested by the author, ranks second-to-last in time but completes the task well. The author feels it now sits just behind Opus 5.5 in capability.

3. Codex + GPT-6 Astra

Codex finishes in 8 minutes, the fastest by far. However, the author observes it "doesn't seem good at thinking," blindly following instructions and even discarding the framework provided by the skill.

4. Claude Code + Opus 5.5

Claude Code is the "undisputed champion" — its video quality is excellent and perfectly explains the problem.

5. Gemini + Gemini 3.8 Flash

Gemini is the "comic relief" with too many flaws.

}