Culpa vs Humanloop

Take the date first, because it decides how to read everything else. Humanloop's own migration guide says the platform closed on 2025-09-08 and that all Files, Versions, Logs, Evaluations and account settings were permanently deleted on that date. Billing stopped on 2025-07-30 and the notice went out on 2025-07-17, so users had 53 days to export through an open-source tool the team published. That's orderly conduct, and it isn't the interesting part. The interesting part is that the docs are still up and they describe a cost model more careful than most live products. Humanloop computed cost, tokens and latency for every Prompt log across all model providers from 2024-04-16. It counted reasoning tokens as their own field rather than folding them into output. It rolled cost up a trace, so an agent run carried the total of its children. And it exposed Cost as an evaluator you could threshold in CI, which is the thing almost nobody ships. Read that list as a specification for whatever you buy next.

• Your replacement tool reports one cost number per call and nothing per agent run.
• Reasoning tokens are folded into output tokens, so a thinking model looks cheap and bills otherwise.
• You can fail a build on latency and you can't fail one on cost.
• Your provider stops returning usage when you stream, and your tool quietly stops counting.
• Nobody can say what next month costs, because the tool that held your history was switched off.

1. Ask your current tool for cost on an agent run, not on a single call.
2. Check whether reasoning tokens are a separate field or folded into output.
3. Try to fail a CI build on a cost threshold, the way Humanloop let you.
4. Stream one call and confirm the token counts still landed.
5. Ask what type money is stored in, and whether totals are computed in it.

Humanloop's published OpenAPI schema types prompt_cost and output_cost as number, format double. A double carries about 15 to 17 significant digits, which is ample for one call and isn't the problem. The problem is summation. Price 40M output tokens on a rate of $0.0012 per 1k, a real shape for a high-volume cheap model, and compare exact decimal against float.

Nobody loses money on that row. You lose the ability to say a total is correct, which is the whole job of a ledger. Culpa stores money as integer pico-dollars and parses rates from decimal strings, so a total is exact by construction rather than by rounding at the end.

The fair reading is that Humanloop built a better cost model than most live tools and pointed it at a different question. Cost was an evaluator there, one signal among faithfulness and latency, judged per version so you could see which prompt got more expensive. That's an engineering question and it's a good one. Culpa asks a finance question with the same numbers. It carries what each customer pays you, so it reports margin rather than spend, it stores money in exact decimal so totals reconcile against an invoice, and it forecasts next month from your own history with a range. The other difference is the one this page is named after. Humanloop's cost history was on Humanloop's infrastructure, and on 2025-09-08 it stopped existing. Culpa's ledger is a database on your own infrastructure, so the number of months you hold is a disk decision.

• Culpa vs Langfuse: An MIT-Licensed Rival, Now Inside ClickHouse
• Culpa vs Traceloop: Spans Measured, Spend Not
• Culpa vs AgentWatch: Enforcement at the Edge, or a Ledger You Own
• Culpa vs Arize: Spans and Gigabytes, or Dollars