Nechával jsem zpracovávat nějaké texty a fable dostal za úkol to jenom orchestrovat a použít opus/sonnet/haiku podle toho, na co jsou nejvhodnější. Haiku rychle vyhodnotil jako ztracený případ, že lepší jsou python skripty (deterministické, auditovatelné). V ostatních to po sobě kontrolovali a i opus udělal semtam chybu; což pak řešil fable aby se už neopakovalo.
I pulled the per-agent token records from all 39 workflow outputs still on disk for this session. The numbers reconcile exactly (per-model sums match the grand total). Orchestrator (Fable 5 main loop) tokens are not metered in these records, and the Python scripts cost zero agent tokens.
ORCHESTRATOR (MAIN LOOP) COST
+------------------+-------+----------+-------------+----------------+-----------+
| Model | Turns | Fresh in | Cache write | Cache read | Output |
+------------------+-------+----------+-------------+----------------+-----------+
| Opus 5 | 1,510 | 2,852 | 11,067,667 | 723,041,429 | 1,771,768 |
| Fable 5 | 1,035 | 1,937 | 14,903,932 | 541,967,725 | 1,283,564 |
+------------------+-------+----------+-------------+----------------+-----------+
| TOTAL | 2,545 | 4,789 | 25,971,599 | 1,265,009,154 | 3,055,332 |
+------------------+-------+----------+-------------+----------------+-----------+
- **Output tokens are the real "thinking" cost: 3.06M total**, of which the Fable 5 stretch (B1 fold through campaign completion) produced 1.28M across 1,035 turns — that covers all the ratifications, seam rulings, gated appliers, source verifications, and commit work.
- **Cache reads dominate raw volume (1.27B)** but that is just the conversation context being re-served from cache on every turn of a very long session - it is the cheapest token \\type by an order of magnitude. Fresh (uncached) input was nearly zero — the prompt cache did its job almost perfectly.
- **Cache writes (26M)** are tool results entering context for the first time: transcripts read for verification, checker reports, screen outputs.
- Put against the subagent fleet: the orchestrators 3.06M output tokens steered 90.85M of subagent work, so orchestration ran at roughly **3.4% of the fleet it directed** — and the Fable 5 half alone, which orchestrated all of Phase B (B1-B6, 21 workflows, ~200 agents), cost 1.28M output, about 1.9% of the ~68M of subagent tokens it dispatched in that stretch.
BY MODEL
+---------------------+-------------+--------+---------------------------------------------------+
| Model | Tokens | Agents | Used as |
+---------------------+-------------+--------+---------------------------------------------------+
| Sonnet 5 | 64,985,636 | 202 | extraction + Phase B drafting, revision passes |
| Opus 5 (1M context) | 20,028,766 | 75 | checkers, confirmers, adjudication, B4, B6 review |
| Opus 5 (standard) | 5,718,443 | 24 | C1 aggregate backfill, wiki name verification |
| Haiku 4.5 | 113,634 | 1 | B0 registry collation (its only appearance) |
+---------------------+-------------+--------+---------------------------------------------------+
| TOTAL | 90,846,479 | 302 | |
+---------------------+-------------+--------+---------------------------------------------------+
BY METHOD
+--------------------------------------------------------------+---------------+-------------+
| Method | Model | Tokens |
+--------------------------------------------------------------+---------------+-------------+
| Phase A extraction drafting (121 episodes, 11 waves) | Sonnet | 41,446,109 |
| Phase A wave checking (2 checkers per wave) | Opus 1M | 5,272,119 |
| C1 aggregate backfill + wiki name verification | Opus std | 5,718,443 |
| B0 setup (arc proposal, registry, wiki) | mixed + Haiku | 673,126 |
| Phase B drafting (B1 pages, B2 slices, B3 layer, B5 hub) | Sonnet | 19,037,108 |
| Phase B revision passes (verify-before-apply) | Sonnet | 4,502,419 |
| Phase B checking, confirmation, proposal adjudication | Opus 1M | 6,013,922 |
| B4 shared-entity continuation (drafting AND checking) | Opus 1M | 6,103,224 |
| B6 independent review + remediation confirms | Opus 1M | 2,737,940 |
+--------------------------------------------------------------+---------------+-------------+
| TOTAL | | 90,846,479 |
+--------------------------------------------------------------+---------------+-------------+
## Notable line items
- The single most expensive method by far was Phase A extraction: each wave ran ~4.2-4.8M tokens (twelve Sonnet drafters reading full transcripts, ~350k each), and its Opus checker pairs cost a steady ~500-575k per wave, about 11% of each wave.
- **Lost work from infrastructure kills: 1,744,221 tokens** with nothing on disk (the wave-7 session-limit kill at 1,397,701 and B1s rate-limit kill at 346,520). Both relaunched fresh; the checkpoint discipline meant nothing else was ever re-paid.
- The B6 reviews four Opus reviewers cost 1,674,392, and the cheapest consequential unit of the whole project was the B3 partition proposal: one Opus agent, 200,247 tokens, whose ratified output steered ~15M tokens of downstream drafting.
- The revision layer (Sonnet, 4.5M) plus the confirm layer (inside the Opus figures, roughly 1.5M) together cost about 6% of the total, and caught the ten wrong checker claims and every silent skip, which is the best ratio in the ledger.