Surriworks · AI usage
Built with AI
Most of these projects were written together with AI coding assistants: Claude Code from Anthropic and Codex from OpenAI. Here is how much text they read and wrote for each project, counted in tokens, as far as the logs on my computer reach.
Counted 14 October 2026
27.3 billiontokens through Claude Code
1.5 billiontokens through Codex
≈ $56,200for the Claude Code part at pay-as-you-go API prices (an estimate)
92%of all of it went into Playfair Rivals
A token is a short piece of text, about three quarters of an English word. B is billion, M is million.
By project
Claude CodeCodex
Claude Code25.32 B≈ $51,100Codex1.16 BClaude (by estimated cost): Opus 5.5 $23,600 · Opus 5 $22,900 · Fable 5.1 $4,300 · Opus 4.8 $109 · Sonnet 5.5 $96 · Sonnet 5 $8
Codex (by tokens): GPT-5.6 Sol 449.5 M · GPT-6 Astra 261.9 M · GPT-5.5 174.3 M · GPT-5.4 156.3 M · GPT-5.6 Terra 66.8 M · other 56.0 M
Claude Code1.6 M≈ $1Codex3.3 MClaude (by estimated cost): Sonnet 5 $1
Codex (by tokens): GPT-5.4 3.3 M
Built with Claude running in Anthropic’s cloud, not on my computer, so none of it is in these logs.
No AI logs for it on this computer.
Other projects
Not on this siteClaude Code1.96 B≈ $5,100Codex354.8 MClaude (by estimated cost): Opus 5.5 $3,700 · Opus 5 $1,000 · Fable 5 $263 · Fable 5.1 $92
Codex (by tokens): GPT-5.4 275.4 M · GPT-6.1 Sol 72.4 M · GPT-6 Astra 6.0 M · GPT-5.6 Sol 914 k
All projects
TotalClaude Code27.28 B≈ $56,200Codex1.52 B
What these numbers miss
- WiFi Sensing is mostly missing. I built it between 25 February and 12 April 2026: 103 commits, 50 of them made with Claude Code (Claude Sonnet 4.6), and 26 chats in a ChatGPT project. Claude Code deletes its local logs after 30 days unless that setting is changed, so those sessions are gone, and ChatGPT shows no token counts for chats. The numbers above hold only what is left: later Claude Code sessions and two Codex sessions. The real total was far bigger.
- Three science projects are not in the logs at all. Neuron Atlas, Latest history of evolution of humans and Tree of Experiments were built with Claude running in Anthropic’s cloud rather than on my computer, so they left no log here.
- Older work can be missing anywhere. The 30-day limit applies to every project, not only WiFi Sensing. The Codex logs on this computer go back to 24 March 2026, my first Codex session on it.
- Most tokens are re-reads. At every step an assistant reads the whole conversation and the files it has open again. These repeats are cached and cost far less than new text: in the Codex logs, 96% of all tokens are cached re-reads and only about 0.4% are text the model wrote.
- The dollar figure is rough. It is Claude Code’s own estimate at pay-as-you-go API prices, and prices for the newest models were not known to it. It is not a bill.
How it was counted
- Claude Code
- Its report from the session logs on my computer, 14 October 2026.
- Codex
- 146 session logs on my computer, 24 March to 10 October 2026, summed per project folder from their token counters.
- WiFi Sensing
- The project’s git history and my ChatGPT project list.
← All sections