Token spend
30 days to 17 September 2026
91.3billion tokens
What one person and a stack of agents actually moves through a month, and what it would have cost at API list rates: $66,325. Cache reads are 97% of everything read, and the stack reads about 262 tokens for every one it writes. Nothing here remembers anything between sessions, so that ratio is what continuity costs.
What on earth am I spending these tokens on? I highlight a selection on /now/.
| Lineage | Tokens | At list rates |
|---|---|---|
| Anthropic | 86.77B | $64,285 |
| OpenAI | 3.17B | $1,943 |
| DeepSeek | 1.35B | $53.62 |
| Z.ai | 18M | $10.24 |
| Moonshot | 14M | $28.36 |
| 12M | $5.05 | |
| not billable | 8,278 | $0.00 |
| unpriced | 0 | $0.00 |
| Model | Lineage | Tokens | At list rates |
|---|---|---|---|
| claude-opus-5 | Anthropic | 38.62B | $29,020 |
| claude-sonnet-5 | Anthropic | 18.54B | $5,351 |
| claude-fable-5-1 | Anthropic | 17.33B | $12,373 |
| claude-fable-5 | Anthropic | 11.93B | $17,208 |
| gpt-5.6-sol | OpenAI | 1.97B | $1,058 |
| deepseek-flash | DeepSeek | 1.23B | $8.88 |
| gpt-6-astra | OpenAI | 429M | $611 |
| codex-auto-review | OpenAI | 401M | $224 |
This is what the work would have billed on the API, not money spent. Calculated and counted from the local records, input, output, cache writes and cache reads, each priced at its own rate:
- every assistant turn in the Claude Code transcripts
- every token_count event in the Codex session rollouts
- every metered call the proxy logged
Chart data: 30 days ending 2026-09-17, as built.