Kimi Code ate 18% of my weekly quota in 3 hours — Here is the log audit comparing it to Claude
Heat trend
The percentage is based on available heat signal, not comment count or independent people.
A developer reported that Kimi Code consumed 18% of their weekly quota in just three hours, raising concerns about its quota calculation compared to Claude Code and Codex. The audit showed Kimi Code processed 10.2M tokens from approximately 30 messages and 110 API requests, resulting in about 90% of a weekly quota being used. In contrast, Claude Code processed 677M tokens from 51 messages and 3,544 API requests on a cheaper plan without similar quota issues, suggesting a discrepancy in Kimi Code's metering.
Is Kimi Code's quota math broken? I compared it with Claude Code and Codex — the numbers don't add up
TL;DR: A single 3-hour session with Kimi Code consumed 18% of my entire weekly quota. On the exact same day, Claude Code processed 66x more tokens on a cheaper subscription without breaking a sweat. Support claims this is "standard product behavior." I ran a forensic audit on the local logs—using Kimi itself to write the parser and measure the data—and the results raise serious questions about how cache tokens are billed.
What happened
I'm an annual subscriber to Kimi Code (Moderato tier). I also use Claude Code and Codex CLI on the same machine for my daily dev workflow.
My Kimi weekly quota kept dying within hours of normal use. When I opened a ticket, support responded:
"All charges are normal. This is standard product behavior based on dialogue turns and historical context."
To see if this was actually "standard," I audited the local session logs across all three agents. I actually had Kimi Code itself write the log parsing script and extract the exact measurement data from the raw session files:
- Kimi: ~/.kimi/sessions/*/wire.jsonl
- Claude: ~/.claude/projects/*/*.jsonl
- Codex: ~/.codex/sessions/*/*.jsonl
Same day, same machine, same user (2026-08-29)
Agent API requests My messages Tokens processed Quota result Kimi Code 110 ~30 10.2M ~90% of WEEKLY quota gone Claude Code 3,544 ~51 677M Fine (cheaper plan)
Over 5 weeks (Jul 24 – Aug 29)
Agent Total tokens processed Claude Code 13.3 BILLION Codex CLI 56M Kimi Code 52M Claude processed 256x more token volume than Kimi over 5 weeks on a cheaper subscription, with much heavier use, and never ran dry.
The Root Cause: Cache Billing & Amplification
Looking at context amplification (how many times the agent re-reads conversation context per turn):
- Kimi: ~10–20x amplification
- Claude: ~300x daily amplification (e.g., Aug 22: 180K new input tokens generated 1.3 Billion processed tokens in agent loops)
Technically, Kimi's agent implementation is more efficient with prompt context than Claude's. So why does Kimi's quota evaporate?
- Unrealistically Small Quota: Kimi's Moderato weekly quota measures out to roughly 11.3M tokens per week (measured: 565,819 tokens = exactly 5% of weekly limit).
- Full-Weight Cache Billing: Kimi appears to bill cache_read tokens at 100% full weight against the subscription allowance. In industry practice, prompt cache reads carry a ~90% discount (~0.1x weight).
Charging cache_read at 1:1 full weight against an 11.3M weekly quota means a standard 3-hour agentic session burns nearly a fifth of your weekly limit just re-reading context.
Support Response
- "Session involved numerous dialogue turns... each request carries full historical context. This is standard product behaviour."