Pay per span. Not per seat.

One timed operation in a trace: a prompt render, a tool call, a retrieval, a completion, an eval. A typical chat request produces 4–8 spans; an agent run with ten tool calls produces 25–40. The SDK counts them for you and the console shows your rate in real time.

On Developer, ingestion is sampled down to stay under 1M; nothing is dropped silently — the console shows the sampling rate. On Team, overage is billed at $8 per million at the end of the month, with an alert at 80% and an optional hard cap.

Eval runs call models, so the provider bills you for those tokens, through your own keys. Tracewell charges nothing per run. The budget setting in tracewell.yaml caps what a single run may spend.

Redaction runs in your process; what you mask never leaves it. Stored payloads are encrypted with per-tenant keys. Enterprise customers can bring their own bucket or self-host entirely, in which case no payload reaches Tracewell at all. We never train on customer data.

No. Point your OTLP exporter at our collector and set two resource attributes. The SDK adds LLM-specific conveniences — automatic token and cost accounting, redaction rules, prompt capture — but everything in the console works on plain OTel spans that follow the GenAI semantic conventions.

Yes, on Team (annual) and Enterprise. We invoice in EUR or USD, net 30. Procurement paperwork — DPA, security questionnaire, SOC 2 report — is available from the console without a sales call.