After finding out recently that Claude Code spent 15-20% of tokens on simply generating the live activity indicators, I took a closer look at Codex cli, since I am a big Codex user.
As it turns out Codex doesn’t commit to the same level of “hidden” token spend. However it does make background/“hidden” model calls, and given that people often complain about token spend and hitting their limits, I thought it would be beneficial for people to become more informed.
Spence, you gotta read what I wrote mate. It’s Claude Code that spends 15-20%. Not Codex. The article is about me investigating where Codex spends tokens, and if it does similar things to Claude. It doesn’t. But still spends tokens on many background tasks.
OpenAI also has hidden dark patterns, token inputs and full-context tool loops for AI to explore, even features that may be incorrectly be used. Things that simply cannot be turned off, regardless of any “personalization” setting, disabling cross-session memory, any other options.
A task, "here’s my file (input tokens), produce a better version of this file (output tokens) will go mad agentic, have tools that can only be described as “try and fail and learn, and call despite uselessness”, and usage and time will be amplified far beyond 20% above an API harness designed for the type of production based on input.
You want the AI reading who your suicide contact is, as the type of cruft always dragged along?
Here’s a dump of web+work skills and plugins with every feature off and already ignoring top-level tools such as web.run, python notebook, and internal code container. It is a summary of what can be called, but where the return type often can’t be determined by AI from the internal presentation. A summary that had to be refined to stay under 200000 characters max of this markdown hosting site:
Be careful when writing “make a web site” because you might see 30 minutes of work and OpenAI published on “sites” feature instead of delivering a text product to you.