2026-07-26 to 2026-08-12 EDT · generated 2026-08-12 11:40
notional API-equivalent across 11 agents over 18 days
These agents run on a Claude subscription, not a metered API key. Every dollar here is a modelled API-equivalent at published list rates, useful for comparing agents against each other. It is not an invoice and nobody billed it.
Where the money goes
Cache reads are the cheapest token there is, at a tenth of the input rate, and they still account for $693 of the total. Cache writes add $395, and everything the agents actually wrote comes to $59.66. Shrinking what an agent carries between turns moves this number; talking less does not.
Daily
Notional cost across all agents, by calendar day.
Per agent, per day
Panels share one y-scale, so height means the same thing in all of them and the quiet agents read as quiet. Switch to the per-agent scale to see the shape of a small agent's week without the loudest one flattening it.
By agent
Ranked by notional cost. The bar is the row, so length and number carry the same fact.
Detail
| Agent | Jobs | Messages | Cache reads | Cache writes | Output | Cost | Per job | Active |
|---|---|---|---|---|---|---|---|---|
| SEO-Ops | 152 | 1,302 | 644M | 15.4M | 1.1M | $487 | $3.20 | 2026-07-26 → 2026-08-11 |
| Mika | 184 | 626 | 206M | 7.4M | 477K | $189 | $1.02 | 2026-07-26 → 2026-08-08 |
| Cindy | 405 | 1,141 | 488M | 21.3M | 272K | $184 | $0.45 | 2026-07-26 → 2026-08-12 |
| Outbound-Ops | 315 | 1,089 | 408M | 4.6M | 373K | $138 | $0.44 | 2026-07-26 → 2026-08-12 |
| PPC-Ops | 72 | 408 | 113M | 5.2M | 337K | $112 | $1.55 | 2026-07-26 → 2026-08-11 |
| Dev-Ops | 39 | 159 | 19.8M | 392K | 89.1K | $14.46 | $0.37 | 2026-08-02 → 2026-08-11 |
| Client-Reporting | 71 | 193 | 25.5M | 773K | 48.2K | $9.37 | $0.13 | 2026-08-02 → 2026-08-11 |
| Rex | 34 | 80 | 6.9M | 278K | 23.9K | $6.31 | $0.19 | 2026-07-26 → 2026-08-09 |
| Nova | 37 | 96 | 9.7M | 394K | 18.2K | $4.46 | $0.12 | 2026-07-26 → 2026-08-11 |
| Aria | 11 | 34 | 3.0M | 256K | 12.3K | $3.24 | $0.29 | 2026-07-26 → 2026-08-09 |
| Spy-Ops | 9 | 37 | 1.9M | 166K | 9.1K | $1.12 | $0.12 | 2026-08-11 → 2026-08-11 |
Models
Synthetic records are Claude Code's own local messages. They carry no model and no charge, so they are counted and priced at zero.
| Model | Messages | Tokens | Cost | Share |
|---|---|---|---|---|
| Opus 4.8 | 2,105 | 714M | $596 | 51.9% |
| Sonnet 5 | 2,047 | 914M | $291 | 25.4% |
| Opus 5 | 826 | 346M | $258 | 22.5% |
| Haiku 4.5 | 105 | 10.3M | $2.44 | 0.2% |
| Synthetic (unbilled) | 82 | 0 | $0.00 | 0.0% |
Method
Claude Code transcripts under
~/.claude/projects/*slock-agents-*, one scope per agent.
Names come from the first line of each agent's MEMORY.md.
A usage record repeats once per content block of the same assistant message. Counting raw lines overstates tokens by roughly 80%, so every message is counted once, keyed on its message id.
Writes are split by TTL because the price differs: a 1-hour write costs 2x the input rate against 1.25x for 5 minutes. This fleet writes 1-hour caches almost exclusively.
Published list rates read from the pricing docs
on 2026-08-11, held in rates.json. Fast mode, US-pinned
inference and batch tier are all applied when a record carries them.