Correction (June 28): the headline number in this post is wrong by roughly 13×. The methodology bug — counting each message once per transcript line instead of once per message — is dissected in the follow-up post. We’ve annotated this post rather than rewriting it, because publishing the mistake honestly is the whole point of this series. Corrected figures appear inline, struck-through next to the originals.
Two days ago, v1.6 shipped cost attribution. There is exactly one honest first test subject for a tool like that: ourselves. So we pointed oc-cost-ops at the full local Claude Code history for this repo — every session, every model call, every cache hit since the repo was extracted — and priced the construction of opchain, with opchain.
This post is the receipts. It’s also, as of June 28, a live demonstration of why oc-signal-forge exists, but we didn’t know that yet when we hit publish.
Method
Claude Code keeps a local transcript of every session as JSONL — one
directory per project, one file per session, one line per event. Each
assistant message carries a usage block: input tokens, output tokens, cache
reads, cache writes, and the model that served it. The report walks every
line, sums the usage, prices it against the current per-model rate card, and
attributes each session to a skill phase by joining timestamps against the
.checkpoints/ history. No sampling, no estimates — the whole ledger,
priced.
If that sounds tidy, know that your own transcript directory is the sketchiest data lake you will ever ETL. The format is undocumented and version-drifted, sessions from abandoned worktrees linger like ghosts, and nothing in the file promises you the invariants you’ll assume anyway. Hold that thought.
The headline number
Building opchain — 27 skills, this site, the Worker, the whole pipeline —
cost $937.41 → $71.73 in API spend.
The corrected number deserves a beat of appreciation: roughly the price of a mid-tier office chair, for a system that plans, builds, audits, ships, and monitors software. The original number implied we’d bought the whole office.
Where it went
The phase attribution (shares as originally computed; the correction shifted them slightly but didn’t change the ordering):
- The build loop dominates — roughly 60% of spend. Generator rounds are the bulk; evaluator rounds run about a third of a generator round and are the best money in the whole budget, because a failed evaluation is a sprint that didn’t ship broken.
- Audits — around 15%.
/oc-auditsweeps are Opus-heavy by design; judgment is the product. - Spec and design — around 10%. Cheap, front-loaded, and the highest leverage per dollar. A spec-phase dollar routinely saves a build-phase ten.
- Everything else — git ops, deploys, checkpoints, orchestration — noise. Single-digit percent. The plumbing phases run on Haiku-class routing and it shows, in a good way.
Caching is the whole ballgame
The single biggest lever in the ledger isn’t the model mix — it’s that oc-claude-api turns prompt caching on by default. Cache reads price at a tenth of fresh input, and an agentic session is mostly re-reading its own context: skill files, specs, the conversation so far. Our hit rates ran high enough that the effective input price for long build sessions was closer to the cache rate than the list rate. If you build with agents and haven’t audited your caching, that — not model choice — is your first dollar.
What we do with the numbers
Attribution without action is a scrapbook. The follow-ups, now that the ledger exists:
- Budget gates per phase, set from measured baselines instead of guesses
— the checkpoint’s
costfield makes an over-budget phase fail loudly. - Model-tier routing review — anywhere Opus is doing Haiku work, the router gets corrected.
- A cost-regression gate beside the eval gate, so a prompt “improvement” that triples spend gets caught in review, not on the invoice.
The seam
We’ll say it before you do: a cost report built by parsing undocumented local transcript files is exactly the kind of artifact that should face hostile review before anyone believes it. We gave it a friendly one. The number above wore a strikethrough within 24 hours, and the follow-up post is the full autopsy — join keys, plausibility checks, and the precise way a confident chart lies.
Run /oc-cost on your own pipeline (install first if needed), or
see the aggregate picture on the dashboard.