What the API actually costs
Measured prompt: 1,559 input tokens for the 112-name table. Output is the
unknown — thinking bills as output and cannot be predicted from the prompt,
so AnthropicClient.usage() records it and turns the estimate into a
measurement after the first real call. Watch output_per_input: above ~2,
thinking dominates the bill and effort is the lever.
Measured, 2026-09-08, one real call at effort=medium: 4,383 input and
1,108 output tokens, $0.0496 per call. That makes a full nine-year backtest
$22.92 and a year of weekly running $2.48.
Two corrections to the estimates that preceded it. Input was 2.8x higher
than a chars/4 heuristic predicted — the schema is billed too, and a dense
numeric table tokenises badly. Output landed at the low end, so
output_per_input is 0.25: thinking does not dominate this call, and
effort is therefore not the cost lever it would be on a harder task. The two
errors happened to cancel.
Running it is nearly free. Scout rebalances weekly — 50 calls a year:
| cadence | calls/yr | @1k out | @2k out | @4k out |
|---|---|---|---|---|
| weekly (Scout today) | 50 | $1.64 | $2.89 | $5.39 |
| daily | 252 | $8.26 | $14.56 | $27.16 |
| three agents mixed | 352 | — | $20.34 | — |
Developing it is where the money goes. One full nine-year backtest is 462 calls ≈ $27. The cache keys on system + user + schema, so any prompt edit re-pays the whole thing:
| iterations | on full history | on a 1-year window |
|---|---|---|
| 5 | $134 | $14 |
| 15 | $401 | $43 |
| 30 | $801 | $87 |
Iterate on a one-year window, then run full history once the prompt is
settled. That single habit is worth ~90% of the development bill. CallBudget
enforces the ceiling in code.