Langfuse vs LangSmith: Which LLM Observability Platform Fits

Choose Langfuse if you need to self-host, want an MIT-licensed core, care about data residency beyond the US and EU, or want unlimited users at a low flat price. Choose LangSmith if your stack is LangChain or LangGraph, or you want agent deployment, sandboxes and Engine’s failure triage alongside tracing. The deciding question is usually deployment: Langfuse self-hosts on any tier; LangSmith self-hosting requires Enterprise.

Yes, as of September 2026. The acquisition was announced 16 January 2026 and Langfuse stated “there are no planned changes to licensing.” The MIT core is unchanged. The change that did land is commercial: supported self-hosted Enterprise is now bundled with a ClickHouse plan, and Langfuse pricing is additive to it.

It depends on your trace shape, and neither headline price answers it. Langfuse bills traces plus observations plus scores — roughly seven units per trace in their own example — at $8 per 100k units over the allowance. LangSmith bills whole traces but publishes no per-trace overage rate, and meters everything else in LCUs at $1.50 and LSUs at $1.00.

It ships one, but it is narrow by design. The Rust ai-gateway/ service relays /openai/v1 and /anthropic/v1 so traces are captured without SDK changes, and the docs state it “does not retry, follow redirects, or accept client routing overrides.” No load balancing, failover, caching or budgets. LangSmith’s is the fuller of the two.

Yes. Both are OTEL-native, so the gateway exports traces to either endpoint, and Langfuse lists TrueFoundry among its gateway integrations. You govern and attribute cost at the gateway; you debug and evaluate in the observability platform.

If you are one developer prototyping, no. Once several teams share provider keys you need budgets, rate limits, per-team attribution and a kill switch — none of which live in a tracing product.