Codex usage percentage vs tokens: why they never move one-to-one
Model weights, cached input rates and parallel sessions shift how fast percentage points drop per thousand tokens.
· 2 min read
Codex usage percentage vs tokens confuses plenty of developers because the CLI talks in tokens inside /status while your account dashboard talks in percentages. Understanding why those two views do not line up 1:1 saves you from thinking a counter is broken when a short prompt moves your percentage faster than a long file read.
Why OpenAI uses percentages instead of token balances
According to UsageAtlas's guide on remaining Codex tokens, Codex allowances are metered in usage units across plans and models rather than a single token bank. A percentage stays stable even if you switch between models or benefit from prompt caching during a multi-turn edit.
No public OpenAI endpoint gives an exact remaining token count for a subscription window. Anything claiming a precise token balance is working backwards from observed sessions.
Four reasons the percentage moves unevenly
| Signal | Why percentage and raw tokens diverge |
|---|---|
| Model changes | Different models consume usage units at different weights per token |
| Cached input | Cached tokens count at OpenAI's cached rate, below full uncached input |
| Short vs weekly windows | A five-hour window can recover while the weekly cap keeps dropping |
| Multi-session work | Multiple open terminals draw from the same account percentage together |
How to translate percentage into rough tokens
When you want a ballpark token budget for planning, measure one typical session: note the percentage before and after, check tokens consumed in /status or UsageAtlas, and divide tokens by points spent. If 12,000 tokens took you from 80% to 76%, each percentage point is about 3,000 tokens for that specific model setup.
Once you change models, such as moving from a lighter model to a heavier reasoning pass, re-measure because tokens per point will drop right away.
Why planning by percentage works better
Tokens are great for comparing which task or repo cost more yesterday. Percentage is what actually stops your session when it hits zero. Setting a threshold alert around 20% inside UsageAtlas gives you one clear nudge before a window runs dry, without checking a web dashboard every hour.
UsageAtlas runs locally on your desktop (free, open source, no account needed for local tracking, and never reads prompts or responses), tracking Codex alongside Claude Code, Cursor, OpenCode, GitHub Copilot and other local assistants.
FAQ
Why did 50k cached tokens barely move my Codex percentage?
Prompt caching bills cached input at OpenAI's cached rate rather than the full uncached rate, so high cached reads move the window percentage less than equivalent fresh tokens.
Should I watch the 5-hour percentage or the weekly percentage?
Watch both. Even if the short five-hour window resets cleanly after lunch, hitting the weekly ceiling still pauses subscription work until the weekly reset.
Can UsageAtlas track both /status tokens and window percentages?
Yes. UsageAtlas reads local session files for token counts and shows supported remaining percentages and reset countdowns in one local view.