Video summary

My Codex Ran 800 Million Tokens in A Day. The Real Story Isn't Cost.

Main summary

Key takeaways

Technology

Summary of the Video

Purpose of the Project (Not Bragging)

The creator built a token burn dashboard to measure how they use AI—not to celebrate a raw number (they mention ~800 million tokens/day).

The real goal is to change how people think about using computers with AI by creating a feedback loop:

  • See token usage
  • Infer behavior quality
  • Improve workflows

Key Idea: “AI Observability” Dashboard

The dashboard is meant to answer questions such as:

  • What are my AI usage habits?
  • Am I using AI effectively?
  • Could I be using it more (and in better ways)?

The emphasis is that dashboards should help you predict where to go next, not just confirm “so far so good.”

Why Codeex Is Central

Codeex is highlighted as the easiest way to measure token usage down to the token, unlike using Claude in certain non-API experiences.

  • With Claude (especially chat/co-work), users may not see exact token counts unless using the API.
  • Because direct metering wasn’t available in their Claude environment, the creator had to do “fancy math” / approximation to infer Claude token usage for the dashboard.

How the Chart Was Created

They used an open-source component: an “Tufty skill” (initially associated with data visualization) to generate a GitHub-style chart.

Using plain-English requirements with Codeex, they produced visuals including:

  • Token burn by day
  • Same-day usage breakdown (what burned tokens vs. what didn’t)
  • Model distribution (including inferred Claude usage)
  • Multiple chart variants hosted on Substack, including comparisons like:
    • Claude vs. ChatGPT vs. Codeex

Feedback Loop + Correlation With Result Quality

The creator argues token burn matters because it is:

  • Easy to measure
  • Correlated with better outcomes

Their framing: higher token spend often means more “delegated intelligence,” increasing the odds of solving problems well.

They also stress not wasting tokens by using mechanisms to:

  • Pause unnecessary automations
  • Slim context windows when they aren’t needed

Concrete Example: Multi-Agent Workflows (Slashworkflows → Claude Code)

They discuss slashworkflows (described as not originally from Codeex) released as part of Opus 4.8, using a workflow approach where:

  • Claude code dynamically creates an orchestration plan
  • It spawns sub-agents to execute multi-step work

They then:

  • Found an open-source skill tied to this concept
  • Ported it to Codeex
  • Used it inside Codeex the same day to complete a large research report (school recommendations) with 3–4 agents

Observed outcomes:

  • More agents → more tokens burned
  • More angles → higher probability of correct/better results

They claim the token chart helped them connect:

behavior change → token burn change → work quality

They also say they couldn’t get comparable flexibility/depth by invoking slashworkflows in other Claude interfaces (like chat/co-work), which is why the chart was especially illuminating.

Chart Customization / Analysis Directives

The creator instructed Codeex (and iterated) to include:

  • Top 10 “AI usage days” with activities (with sensitive info scrubbed)
  • Token usage trend logging over time
  • A logarithmic axis to handle massive scale differences (millions to near a billion tokens/day)
  • Color contrast/readability improvements
  • Multiple model views

They also mention Codeex provided a quiz/claim-check mechanism for inferring Claude usage, yielding a tight range approximation.

Deployment / Automation of the Dashboard

They claim Codeex handled deployment details including:

  • DNS/domain changes
  • Hosting at a domain like token.burn.mmarkdown (as stated)

Other “Observability” Uses Beyond Dashboards

They also use Codeex to improve day-to-day AI productivity by automating organization, including:

  • Automatically labeling and organizing screenshots/files
  • Handling workflows so the user doesn’t need to maintain folder structures manually
  • Additional optimization, such as:
    • Email/Slack handling
    • A “chief of staff” thread that spawns sub-agents for project work while keeping contexts clean

Broader Argument: Transparency + Community Accountability

They encourage sharing token burn charts publicly to create a learning loop across users.

They also frame it as professional transparency, where prospective employers might compare AI usage intensity (example: “burning 3 million tokens/day” vs. far higher).

They note Codeex adoption is still early (they mention ~0.6%), emphasizing that this is still exploratory.

Interpretation of Model Behavior

The creator argues models are “grown, not made” and aren’t fully understandable like classic software. They criticize:

  • Journalism for misrepresenting AI as traditional software

They reference transformer architecture and mention separate videos on the topic.

Main Speakers / Sources

  • Speaker/source: Nate (referred to as “Nate” throughout, likely the video creator)
  • Primary tools mentioned:
    • Codeex
    • Claude (including Claude Chat/Co-work and Claude API)
    • Opus 4.8
  • Skills / third-party components mentioned:
    • Open-source Tufty skill
    • Open-source slashworkflows skill/port

Original video