Why we bill on processed bytes, not trace count

Most observability tools bill per event. It sounds fair until you realize what it does to your behavior: every span you add, every nested tool call you capture, every extra bit of context you attach makes your bill go up — even when those traces are tiny. Per-event pricing quietly teaches you to instrument less, which is exactly backwards.

Pricing that tracks reality

A trace for a one-line classification and a trace for a long agent loop with twelve tool calls are not the same thing, and they should not cost the same thing. Currai bills on the data it actually processes:

• Usage is metered on processed bytes, not on trace or span count.
• Retention follows your plan — keep data as long as your plan allows.
• A small trace costs little; a large one costs proportionally more.
• Cost is predictable because it tracks real volume, not row counts.

# retention follows your plan; billing follows processed bytes
trace = currai.trace(name="chat-turn", metadata={"retention_days": 30})
trace.generation(name="openai.chat", model="gpt-4o-mini", input=messages).end(output=reply)

Instrument freely

The point of this model is to remove the tax on good instrumentation. Add the span. Capture the tool call. Attach the metadata. The richer your traces, the better your debugging — and you only pay for the data that richness actually represents.