Once you ship an LLM feature you need to see what it's actually doing — traces, token costs, latency, eval scores. Three tools dominate, and they price on very different units.
| Free tier | Paid entry | Meters on | Self-host | |
|---|---|---|---|---|
| Langfuse | 50k units/mo | Core $29/mo (100k units, unlimited users) | usage (units), seats free | yes — MIT, free |
| LangSmith | 10k traces/mo | Plus $39/seat/mo + $0.50 per 1k traces | seats + traces | limited (enterprise) |
| Helicone | 50k requests/mo | Pro $20/mo | requests | yes — open source |
The cost gap at volume is enormous. At roughly 1 million events a month, Langfuse Core works out to about $100 (the $29 base plus usage increments). The same volume on LangSmith Plus is around $2,500, because LangSmith charges per seat and per trace — every engineer you add raises the bill, and so does every trace.
Langfuse is the value pick and the one to self-host if you want to: MIT-licensed core, usage-based pricing that keeps team growth free, S3-compatible archival for old traces. It's become the default for teams not married to the LangChain framework.
LangSmith makes the most sense if you're already building on LangChain/LangGraph — tracing is automatic, the prompt playground and eval tooling are tightly integrated, and for a small team the seat cost is tolerable. It gets expensive fast as you scale.
Helicone is the lightweight option: it's a proxy, so integration is a one-line base-URL change, and its strength is cost analytics and caching (which can shave 20–40% off your model spend). Less depth on evals and complex agent traces.
Framework-agnostic and cost-sensitive: Langfuse. All-in on LangChain: LangSmith. Want a proxy that also cuts your bill: Helicone.