Skip to content
Notifications
Clear all

Beginner question: what's a 'generation' vs an 'observation'?

1 Posts
1 Users
0 Reactions
1 Views
(@cloud_cost_auditor)
Estimable Member
Joined: 3 months ago
Posts: 109
Topic starter   [#21543]

Alright, I've been spelunking through the Langfuse docs and pricing page, and I see they charge per "generation." They also keep talking about "observations." As someone who has been burned by nebulous cloud pricing before, I need to understand what I'm actually going to be billed for.

Can someone break down the *practical, cost-related* difference between a 'generation' and an 'observation'?

From what I gather:
* An **observation** seems to be the basic unit of anything you log—a trace, a span, an event. It's the generic container.
* A **generation** is a specific type of observation for LLM calls.

My immediate questions are:
* Does every single call to an LLM provider (OpenAI, Anthropic, etc.) count as one generation, regardless of token count or cost?
* What about simple tool calls or non-LLM spans logged as observations—are those free in the pricing model, or do they just fall under a different limit?
* If I have a trace with 5 LLM calls and 10 other events/spans, is that 5 generations and 15 total observations?

I'm trying to map this to real usage. If my app makes 50,000 LLM calls in a month, that's 50,000 generations, putting me in the $199/month tier. But are the other 200,000 non-LLM observations I might log going to trigger an overage or are they just... included?

A concrete example would help before I even consider a trial.


Show me the bill


   
Quote