Benchmarks
Cost per feature: hybrid vs pure cloud
The claim Zecor is built around: most landable work never needs a frontier model, so the metered cost of a shipped feature is a fraction of a cloud-only agent's.
Method
Each run works an identical backlog against the same repository. Recorded per task:
- which tier landed it (T0 codemod, T1 local, T3/T4 cloud)
- metered input/output tokens and their USD cost
- wall-clock from dispatch to a green gate
- escalations before it landed
Two configurations: hybrid (the default ladder — T0/T1 first, cloud only on escalation) and cloud-only (every task goes straight to a frontier applier). Same model versions, same gate, same backlog.
Results
Pending the first full runs. Zecor records every attempt in a local ledger; these figures are published once that ledger has representative data rather than estimated. The table below shows the shape of what is measured.
| Metric | Hybrid | Cloud-only |
|---|---|---|
| Tasks landed without a frontier call | — | 0 |
| Median metered $ / landed feature | — | — |
| Median wall-clock / landed feature | — | — |
| Escalations / task (mean) | — | n/a |
Reproduce it
uv run zecor run --until 07:00 uv run zecor metrics # land rate by tier, cycle time, est. cost
zecor metrics reads the same ledger and reports land rate by tier,
cycle time, and an estimated cost. It states plainly when the ledger has no
runs yet.