Cost Rollup
By Yad Konrad — @0bserver07
Estimated at Opus 4.x public pricing (May 2026): input $15/M, output $75/M, cache_read $1.50/M, cache_write_5m $18.75/M, cache_write_1h $30/M.
Cost composition
pie showData
title $ share by token pool (total $3,878.59)
"input" : 3.03
"output" : 793.63
"cache_read" : 1596.30
"cache_write_5m" : 77.02
"cache_write_1h" : 1408.62
Two pools combine to 77.5% of the bill — cache_read (41%) and cache_write_1h (36%). Output is third at 20.5%. Raw input tokens are negligible because the system/tool prompt was almost always cached.
By pool (orchestrator + 58 workers)
| Pool | Tokens | $/M | Cost | Share |
|---|---|---|---|---|
| input | 202,129 | $15.00 | $3.03 | 0.1% |
| output | 10,581,714 | $75.00 | $793.63 | 20.5% |
| cache_read | 1,064,199,056 | $1.50 | $1,596.30 | 41.2% |
| cache_write_5m | 4,107,469 | $18.75 | $77.02 | 2.0% |
| cache_write_1h | 46,953,891 | $30.00 | $1,408.62 | 36.3% |
| total | 1,126,044,259 | — | $3,878.59 | 100.0% |
Cache reads are the bulk of tokens but cheap per-token. Output is the conventional cost driver, but on long orchestration sessions like this one, cache_write_1h can match or exceed it.
By role
| Role | Sessions | Tokens (total) | Cost | $/session |
|---|---|---|---|---|
| orchestrator | 1 | 487,710,910 | $1,283.73 | $1,283.73 |
| workers | 58 | 638,333,349 | $2,594.86 | $44.74 |
| total | 59 | 1,126,044,259 | $3,878.59 | — |
Orchestrator carries ~33% of total cost despite being a single session, because it holds the long context (full project + tool list + every dispatch result) and gets recomputed many times.
By wave
| Wave | Slug | Workers | Total turns | Cost | $/worker |
|---|---|---|---|---|---|
| 0 | sanity | 1 | 98 | $56.18 | $56.18 |
| 1 | search | 6 | 531 | $213.94 | $35.66 |
| 2 | local-rules | 5 | 503 | $210.31 | $42.06 |
| 3 | rl-hidden-state | 5 | 594 | $273.71 | $54.74 |
| 4 | history-fastweights | 5 | 496 | $215.84 | $43.17 |
| 5 | predictability | 4 | 382 | $159.86 | $39.96 |
| 6 | lstm-1 | 6 | 561 | $202.65 | $33.78 |
| 7 | lstm-2 | 5 | 600 | $254.87 | $50.97 |
| 8 | evolutionary | 4 | 572 | $257.95 | $64.49 |
| 9 | deep-mlps | 4 | 392 | $146.34 | $36.59 |
| 10 | modern | 5 | 642 | $286.19 | $57.24 |
| 11 | v1.5 | 8 | 868 | $317.03 | $39.63 |
| — | workers total | 58 | 6239 | $2,594.86 | — |
Wave 11 (v1.5, 8 heavyweight stubs) and wave 10 (modern, 5 stubs) topped the per-wave cost. Wave 3 (RL hidden state) was unexpectedly expensive per worker — partial-observability environments take more turns to get right.
Per-wave cost (bar)
xychart-beta
title "Worker cost by wave (USD)"
x-axis "Wave" [W0, W1, W2, W3, W4, W5, W6, W7, W8, W9, W10, W11]
y-axis "$ cost" 0 --> 380
bar [56, 214, 210, 274, 216, 160, 203, 255, 258, 146, 286, 317]
Per-worker cost distribution
- Min: $20.77 (
pole-balance-markov-vac) - Median: $41.05
- Mean: $44.74
- Max: $122.05 (
pipe-6-bit-parity)
Workers built simple stubs in ~$25, complex stubs in ~$90+. The orchestrator is its own outlier at ~$1,283.