I ran two tasks on small repositories:
1. Recommend movies based on an existing movie database. 2. Analyze a website's codebase.
Neither task completed. The entire 5-hour quota was exhausted within minutes.
I analyzed the actual session logs, including background agents:
- Fresh input: 368,203 tokens - Cached input: 1,183,744 tokens - Output: 39,324 tokens - Total: 1.59M tokens
At Kimi's published API rates, that's just $1.37 worth of usage before hitting the 5-hour limit which is easily less than 5% of what Claude Code offers at $20 subscription.
The logs show: "429 You've reached your 5-hour usage limit."
For comparison, Claude Code on a $20 subscription processed roughly 381M tokens over 4-5 hours using Opus 5.5 xhigh, mostly cached input. That's not a controlled benchmark, but the difference in usable capacity is staggering.
Both Kimi tasks failed without delivering results. I have video evidence and detailed logs, but won't publish private repository data.
Video: https://www.youtube.com/watch?v=hSr-DLn88AM
Is this a quota-accounting bug, or are Kimi's $19 subscription limits intentionally this restrictive?
Disclosure: I also develop a competing coding agent. These tests used Kimi's official CLI.
3 comments