Langfuse stores the trace. CostSlice stores the cost.

Langfuse is LLM observability. It stores the trace: the prompt, the span, the eval, and often a cost number on that span. Teams use it to see what the model did and whether the answer was any good. That is a real job. CostSlice is not that job.

CostSlice stores the cost. Not the trace. Prompt and completion bodies are not written to disk. We keep model, tokens, latency, status, timestamps, and the tags you send. The screen is a customer P&L, not a session replay.

Langfuse can show cost. The unit is the span. Finance cannot raise a price from a span. They need this customer, this month, OpenAI cost versus plan price. That is OpenAI cost P&L. Observability does not become a P&L because a dollar sign appeared on a trace.

We do not replace Langfuse. If you need traces and evals, keep Langfuse. If security will not let you store what people typed, you still need a meter. That meter is CostSlice. Some teams run both. The trace stays in Langfuse. The invoice question stays here.

Tag the tenant on the CostSlice request. x-cs-tenant is an opaque id, never an email. x-cs-feature is the SKU. x-cs-env is prod or staging. Keep the OpenAI API key in your environment. Point base_url at https://api.costslice.com/v1. Send x-cs-key.

Helicone is closer to a gateway that logs. If the comparison you wanted was logs and routing, read vs Helicone. The privacy page for the meter is without storing prompts.

Missing tags do not drop the completion. Untagged still returns. We do not cap a tenant because a P&L row is red. Starter is a hard cap on the CostSlice plan. That is not an observability sampling rule.

Claude is not on CostSlice today. OpenAI only. Langfuse can sit on many models. That is their product. We do not pretend to ingest those traces.

A security review will ask two questions. Does the meter keep text? No. Does the observability tool keep text? Usually yes — that is the trace. Answer them separately. Do not collapse the meter into the trace because both show a dollar sign. Cost is a row. The trace is a session. Finance signs the row. Engineering debugs the session.

Export CSV when you need the allocation in a sheet. Retention follows the plan: seven days on Starter, 90 on Pro, a year on Scale. Langfuse retention is their setting. Ours is the P&L window. The sample row acme-health is −101%. That number is cost versus plan, not a span duration.

Starter $0 / Pro $79 / Scale $199, overage $6/100k. No credit card on Starter. Pricing on the homepage wins if a number here is stale.