Arize Phoenix vs LangSmith
If you are building an LLM app and want to see what your traces, prompts, and evals are doing, both of these will get you there. Arize Phoenix is open-source tracing you run yourself, built on OpenTelemetry. LangSmith is a hosted platform from the LangChain team that you sign into and send data to.
| Dimension | Arize Phoenix | LangSmith |
|---|---|---|
| What it is | Open-source LLM tracing and eval tool from Arize | Hosted tracing and eval platform from LangChain |
| License / model | Open source | Commercial / hosted, with a self-host option |
| Where it runs | Self-host or local, on your infra | Mostly their cloud |
| Instrumentation | OpenTelemetry, framework-agnostic | Native LangChain and LangGraph, plus an SDK for other stacks |
| Data control | Stays where you put it | Goes to their platform by default |
| Evals | Built-in eval library, code-first | Datasets and evals with a UI to run them |
| Team features | Lighter, engineer-focused | Stronger sharing, dashboards, and prompt management |
| Best for | Self-hosting and any-stack tracing | LangChain teams who want a managed platform |
You want to own the data and run tracing on your own infra, or you are not on LangChain and want a vendor-neutral, OpenTelemetry-based setup. Phoenix is also the easy answer when you just want to spin something up locally and poke at traces without signing up for anything.
Your app is built on LangChain or LangGraph and you want tracing, datasets, and evals working out of the box without running a service yourself. LangSmith pays off most when a team needs shared dashboards, prompt management, and a hosted place to collaborate on evals.
If you are already deep in LangChain, LangSmith is the path of least resistance and I would just use it. If you are not, or you care about self-hosting and keeping trace data in-house, Phoenix is the cleaner fit and costs nothing to try. The common mistake is picking LangSmith by reflex because you used LangChain once, then paying for a hosted platform when a self-hosted open-source tool would have done the job. Both give you real trace visibility. Decide on who owns the data and whether you want to run the service yourself.