Langfuse vs Helicone vs LangSmith - which LLM tracing tool?

What is the duel about?

As soon as an LLM workflow grows beyond PoC status in May 2026, the same question surfaces: how many tokens did this request consume, why did the model hallucinate here, which prompt version was active at the time? Without a tracing tool the answer is "we do not know". Three platforms hold the market top: Langfuse (Berlin startup, MIT, EU cloud), Helicone (San Francisco, Apache 2.0, proxy model) and LangSmith (LangChain Inc., proprietary, US-only cloud).

All three answer the same core questions - per-request trace view, cost and latency metrics, prompt versioning, evaluation datasets. The differences sit on three axes that directly hit Swiss fiduciary and law-office setups.

First: hosting model. Langfuse offers MIT OSS with all core features in self-host plus an EU cloud in Frankfurt. Helicone is Apache 2.0, self-host possible, plus US and EU clouds. LangSmith is proprietary, primarily US cloud, self-hosting only on the enterprise tier from five-figure annual costs.

Second: integration model. Helicone captures traces as an HTTP proxy - one URL swap in code configuration suffices, no SDK code. Langfuse and LangSmith work via SDK wrappers or OpenTelemetry. The proxy approach is faster to set up, but it requires that all LLM traffic flows through the Helicone endpoint.

Third: framework binding. LangSmith is built for LangChain/LangGraph stacks - deep hooks, automatic tracing without configuration. Langfuse is framework-agnostic via OpenTelemetry - LangChain, LlamaIndex, Haystack, raw OpenAI calls. Helicone is framework-agnostic as well.

Why the choice matters

Three hard factors make the tool choice in May 2026 a question of trust, not taste.

Data protection (revDSG and GDPR): a fiduciary office processing client data through an LLM application sends every tracing detail to the observability platform - question prompts, answers, sometimes tool-call arguments with cleartext names, receipts, tax figures. If that traffic lands on a US server, you are transferring data to a third country - Transfer Impact Assessment required, standard contractual clauses, EDSB risk. Langfuse with the Frankfurt cloud or self-host solves that, Helicone EU cloud as well. LangSmith primarily on US cloud remains uncomfortable for DSG-compliant setups.

Setup speed: Helicone wins clearly - swap the URL, done. Realistically two to five minutes from account creation to the first trace in the dashboard. Langfuse via OpenTelemetry or a Python/JS decorator takes 15-30 minutes to set up but produces a richer trace tree. LangSmith in the LangChain stack is essentially automatic, but only in that exact stack.

Feature depth for evaluation and prompt versioning: Langfuse clearly leads here in May 2026. Built-in eval datasets, LLM-as-judge pipelines, prompt versioning with A/B tests, automatic cost aggregation per user/session/feature. Helicone covers the same areas but is a notch lighter on eval and prompt versioning. LangSmith also has strong eval tools, bound to LangChain data types - convenient for LangGraph teams, less universal.

The three platforms in detail

Langfuse (MIT, Berlin/Frankfurt, EU cloud + self-host): the OSS market leader for LLM observability in May 2026. Architecture over the OpenTelemetry span model - every LLM call is a span, tool calls and retrieval steps are child spans, a RAG workflow becomes a tree-structured trace. Cloud in Frankfurt (eu-central-1) plus self-host via Docker Compose or Helm. Prompt management with versioning and A/B tests, eval datasets, LLM-as-judge with configurable prompts, score aggregation, user-session tracking. SDK for Python and JS/TS, plus an OpenTelemetry endpoint for any other language. Cloud pricing: free for 50k events/month, Pro from USD 29/month. Self-host completely free.

Helicone (Apache 2.0, San Francisco, US+EU cloud + self-host): proxy-centric model as the core differentiator. Instead of an SDK wrapper, you change the base URL of the LLM client from api.openai.com to oai.helicone.ai, add an auth header, and Helicone intercepts every request. Setup in under five minutes, no code refactor needed. Second option: async logging via SDK for teams who dislike a proxy. Features: cost tracking, latency metrics, caching, rate limiting, an own playground. Eval and prompt versioning are present as of May 2026 but lighter than in Langfuse. EU cloud enables DSG-compliant setups, self-host for maximum control.

LangSmith (proprietary, US cloud, LangChain Inc.): the in-house observability tool of the LangChain project. Inside LangChain or LangGraph code a single environment variable enables automatic tracing - every chain step appears as a span. Best eval integration for LangChain data types, hub for prompt sharing, integrated deployment pipeline. As of May 2026 still primarily US cloud; self-hosting only in the enterprise tier (typically from USD 30000+/year). For DSG-strict setups in Switzerland that is the most critical hurdle.

Tool selection in 6 steps

1. 01Check data sensitivity: client data in prompts? If yes: EU hosting required (Langfuse EU or Helicone EU or self-host).
2. 02Identify framework stack: pure LangChain = LangSmith comfort advantage; LlamaIndex / mix = Langfuse; framework-agnostic = Helicone.
3. 03Budget setup time: 5 minutes = Helicone (proxy); 30 minutes = Langfuse SDK; automatic = LangSmith inside LangChain code.
4. 04Clarify feature requirements: prompt versioning A/B + LLM-as-judge = Langfuse leads; simple cost tracking = all three.
5. 05Decide the self-host question: EU cloud is enough = Langfuse Frankfurt cloud or Helicone EU; own rack = prefer Langfuse self-host.
6. 06PoC with two weeks of data: hands-on first, then commit. Estimate trace volume, project costs, draft a production plan.

Recommendation by scenario

Swiss fiduciary or law office with DSG obligation, client data in traces: Langfuse with the EU Frankfurt cloud or self-host on Hetzner. The standard choice in May 2026 when data protection and audit-readiness count. Self-host runs on an AX41 server with Postgres and ClickHouse, monthly operating cost below CHF 60.

5-minute setup, fast insight without code change: Helicone. Anyone needing to see in a PoC "where do my tokens go, what latency does OpenAI deliver today, which client drives 80 percent of cost" sets Helicone up in five minutes and has the answer from minute six. Choose the EU cloud if the PoC is meant to grow into production.

Team fully in the LangChain/LangGraph stack, US cloud accepted: LangSmith. LangChain-native integration saves setup time, the eval tooling fits LangChain data types, the prompt hub is solid. Precondition: no DSG risk or SCC plus TIA documented.

Multi-framework setup (LangChain + LlamaIndex + raw calls) mixed: Langfuse. OpenTelemetry-based, all three sources flood the same dashboard, unified trace view. As of May 2026 the only platform among the three that handles this mix elegantly.

Self-host as a hard requirement (law firm under professional secrecy): Langfuse or Helicone. Both are OSS, both run in your own rack. Langfuse feels more mature in self-host in May 2026 (clear Helm charts, regular releases). Helicone runs too, but its self-host path is better documented for cloud setup.

Hybrid: some apps in LangChain, others not: Langfuse for the non-LangChain apps, LangSmith optional for LangChain - or Langfuse for everything. Double-tooling rarely pays off.

When none of the three fits

If LLM usage in May 2026 stays below 1000 requests per month and is a PoC due to disappear in four weeks, every one of the three tools is overkill. Simple logfile entries plus an Excel evaluation at month end suffice.

If the primary need is not LLM observability but application performance monitoring (database latency, HTTP statuses, app memory usage), reach for classic APM tools such as Grafana, Datadog or Sentry. Langfuse and Helicone see the LLM call well but they do not see that your Postgres index is missing.

If the compliance posture is so strict that no external software may observe client data at all - some law firms interpret SCC Art. 321 that way - only self-host plus an own code-base review remains. Langfuse self-host is the most mature path in May 2026, Helicone self-host the easier alternative.

If the LLM setup works exclusively with local models (Ollama, vLLM, llama.cpp) and never calls a cloud API, the added value of a tracing tool is limited - there is no per-token cost tracking at stake. In that case an own logger plus a Grafana dashboard, OpenTelemetry directly into Tempo or Jaeger, often suffices.

Trade-offs

FAQ

Related topics

Sources

1. Langfuse - official documentation · 2026-05
2. Helicone - official documentation · 2026-05
3. LangSmith - official documentation · 2026-05
4. OpenTelemetry GenAI semantic conventions · 2026-04