Replay
Full prior transcripts accumulate inside every later turn. Summing a single average turn misses that growth.
ORCHESTRATION FINOPS
Price the prompt your harness assembles.
Compare full-history replay with a byte-stable cached prefix and bounded compact history. The curve shows what model choice cannot fix: tokens your code keeps sending again.
TURN TOKEN CSV0 turns
STABLE PREFIX4,000 tokens
HISTORY CAP4,500 tokens
PRICES / 1M$3 · $15 · $0.30
Synthetic six-turn trace ready. Replace prices with measured values.
Cost substrate
Compute the trace to reveal replay and cache-shaped growth.
baseline cost$0.0000
harness cost$0.0000
observed savings0.0%
raw CPM0 → 0
full replaycache-shaped
0 plotted turnsBASELINE INPUT0
Stable prefix + current user/retrieval + every prior transcript.
HARNESS INPUT0
0 cached + 0 volatile
Quality and latency are not inferred; this tool prices only the supplied token trace.
The cheapest token is the stable one your harness stops rebuilding.
Shape
A stable prefix earns cache-read pricing. Compaction caps the volatile history; retrieval and current user tokens remain full-price.
Measure
Replace the fixture with production traces. Cache eligibility, quality, latency, provider rules, and offload overhead still require live measurement.