LangSmith vs Helicone vs Phoenix vs Langfuse: The 2026 LLM Observability Buyer's Guide

Pick LangSmith if your stack is built on LangChain or LangGraph and you want the tightest agent-tracing UX inside that ecosystem. Pick Langfuse if you want true open-source (MIT) with full self-host parity and you are comfortable with ClickHouse as the new owner. Pick Arize Phoenix if you are committed to OpenTelemetry and want a framework-neutral tracing layer, especially for LlamaIndex or DSPy workloads. Pick Helicone if you need zero-code instrumentation today and can accept that the product is now in maintenance mode following the Mintlify acquisition.

LangSmith is LangChain's proprietary observability product, deeply coupled to the LangChain and LangGraph ecosystem, with SaaS pricing starting at $39/seat/month on the Plus tier. Langfuse is the open-source (MIT) alternative — framework-agnostic, with full self-host parity to its cloud version, and now backed by ClickHouse following the January 2026 acquisition. The most practical difference is what happens when you change frameworks: LangSmith becomes awkward outside LangChain; Langfuse stays useful. The most practical similarity is that both produce excellent trace visualization for agent reasoning chains.

For teams deep on LangChain, no — LangSmith's native integration is hard to beat. For teams that use multiple frameworks (LangChain alongside LlamaIndex, custom OpenAI SDK code, LiteLLM, or vendor SDKs), yes — Langfuse is usually the better long-term bet because of its framework-agnostic design, MIT license, and full self-hostable parity. The ClickHouse acquisition adds operational confidence; the Hobby tier (free, 50k events/month) and Core tier ($29/month) make it accessible at the smallest team sizes.

Phoenix is built around ML observability broadly, not just LLMs. It covers LLM traces, retrieval quality (critical for RAG agents), embedding drift, and model performance monitoring in one place. The OpenInference instrumentors give it first-class coverage of LangChain, LlamaIndex, and DSPy without forcing you to commit to any one framework. Phoenix Evals is the most mature open evaluation library, which matters once you move from "is the agent working?" to "is the agent improving?" For teams running both traditional ML and LLM agents in the same stack, Phoenix is the natural choice.

Helicone is an HTTP proxy, not an SDK. You change the base URL in your environment and every LLM call gets logged automatically — no instrumentation code, no decorators, no SDK wrappers. That makes it the fastest tool in the category to deploy. The trade-off is visibility depth: because it works at the HTTP layer, it sees requests and responses but has less native context on agent reasoning chains than tools that use SDK instrumentation. Best for teams that need immediate cost visibility on raw API calls with minimal setup. Worst for teams that need deep agent-step tracing.

LangSmith: Free Developer (5k traces/month, 14-day retention); Plus $39/seat/month + $2.50 per 1k overage; Enterprise custom. Helicone: Free 10k requests/month; Pro $79/month; Team $799/month; Enterprise custom. Arize Phoenix: self-host free (Elastic License 2.0); Arize AX cloud has Free 25k spans/month, Pro $50/month for 50k spans, Enterprise custom. Langfuse: Hobby free (50k units/month); Core $29/month; Pro $199/month; Enterprise $2,499/month; overage $8 per 100k units. Self-host is free across Phoenix and Langfuse; the cloud SaaS pricing is the more comparable axis for most teams.

For prebuilt evaluation templates, Arize Phoenix Evals is the strongest open library in the category, with reference implementations for hallucination, relevance, toxicity, faithfulness, and RAG-specific metrics. Langfuse Datasets v2 (shipped 2026 alongside the ClickHouse integration) is closing the gap and is the strongest option for teams that want eval workflows tightly integrated with the same tool they use for tracing. LangSmith's eval library is good inside LangChain but does not extend as cleanly to non-LangChain code. Helicone is not the right pick if eval depth is a priority.

Self-host when data sovereignty, regulatory compliance, or token-level data sensitivity rules out a third-party cloud. Langfuse (MIT) and Arize Phoenix (Elastic License 2.0) are the credible self-host options; both ship full feature parity to their cloud versions. Helicone is self-hostable via Helm but feature investment has slowed post-acquisition. LangSmith self-host exists on Enterprise tier only and is closed-source. Cloud is the right answer for most early-stage teams: faster setup, no infrastructure overhead, and the free tiers on all four tools are generous enough for evaluation and small production workloads.

All four track token spend per call, per session, and per user. The meaningful differences are in alerting and aggregation. Helicone's proxy model makes per-route cost visibility trivial. Langfuse has the strongest cost-per-experiment view (helpful when iterating prompts). Phoenix surfaces cost in the same plane as quality metrics, which is useful when trade-offs are explicit. LangSmith's cost tracking is good inside LangChain runs but less granular for hybrid stacks. For preventing a runaway retry loop from burning your monthly budget overnight, Helicone's alerting is the most mature; Langfuse v3 closed much of the gap in 2026.