Cost and budgets

Culprit runs on your keys, so you pay your providers directly. There's no Culprit charge.

What costs money#

• Your agent's runs. Every task run at every tested commit calls your model. This is most of the cost.
• Culprit's own model calls, a handful per run: ranking hunks, proposing a fix, generating audit mutants. A bisect alone makes no model calls of its own.

How much#

It depends on your suite size, k, your model and how many commits are in the range. Measured on our test agent (26 tasks, k = 3, 6 planted regressions, Culprit 0.1):

On GLM 5.3, staged baselines cut a bisect from 190 agent runs to 98 on average, and the cost by 44%, with the same accuracy.

Budgets#

Every command that spends money takes --budget-usd, a hard cap for that command. Each command's page lists its default.

When a run reaches its cap, it stops with budget_exceeded and reports the narrowest range found so far. Nothing is lost: continue it with a higher cap using culprit resume.

The ledger-wide cap#

Each ledger also has a $75 lifetime safety cap across all its runs, so a script that loops can't run away with your credit. When it's reached, every command stops with "ledger total reached the $75.00 global cap". To carry on, start a new ledger, keeping the old one for its reports:

mv .culprit/ledger.sqlite .culprit/ledger-1.sqlite

or pass --db path/to/new.sqlite to each command.

Spending less#

• Use a cheaper victim model for bisecting. In our tests, accuracy held on every model while cost varied tenfold. See models.
• Narrow the range. A good ref closer to the bad one means fewer steps.
• Keep k at 3 unless your agent is very noisy.
• Let baselines be reused. Repeated runs over the same ends reuse earlier measurements.
• Run culprit export-costs to see per-run cost, agent runs and time for everything in the ledger.