LangSmith vs Langfuse vs Arize Phoenix: LLM Observability

LangSmith, Langfuse and Arize Phoenix compared on tracing, evals, self-hosting and pricing, with a pick for each type of team.

Langfuse is the safest default for most teams. It is open source and free to self-host.

Short answer

On the scored.tools score, Langfuse leads this comparison at 8/10, ahead of Arize Phoenix (7.2/10) and LangSmith (6.8/10). Our verdict on Langfuse: A strong default for teams that want LLM tracing and evals with an open source, self-hostable base.

As of . Scores and prices come from the tool pages linked below; nothing a vendor pays for changes a score.

At a glance

As of . scored.tools scores last updated . Prices are as listed on each vendor's site when we last updated that tool, dated per row.

Toolscored.tools scorePlans and listed pricesPrices dated
LangSmith6.8/10
  • Free: Free
  • Plus: $39 per seat/month
Langfuse8.0/10
  • Hobby: Free
  • Core: $29/month
  • Pro: $199/month
  • Enterprise: $2499/month
Arize Phoenix7.2/10
  • AX Free: $0/month
  • AX Pro: $50/month
  • Startup: Startup pricing available
  • AX Enterprise: Custom

Why this comparison matters

Once an LLM app moves past a demo, you need to see what happened inside each request. That means which prompt went out, which tool the agent called, what the retriever returned, how long each step took and what it cost. Logging raw completions to stdout stops working at the first multi-step agent. Observability platforms fill that gap. They record each run as a tree of spans and let you score outputs with evals, and the better ones let you turn bad traces into test datasets.

Three names come up in nearly every shortlist. LangSmith is LangChain's hosted platform for tracing, evaluating and deploying agents. Langfuse is an open source platform you can self-host or use as a cloud service. Arize Phoenix is Arize's open source, local-first tracing and eval tool, with a managed tier called Arize AX on top.

All three cover the basics. They differ on licence, on how they charge, and on how tightly they assume you use one framework. Those differences decide which one fits. This piece compares them using vendor pricing pages, public GitHub data and the scores in our tool database. scored.tools did not run these tools hands-on, so claims about behaviour are attributed to their docs or to named sources.

Scores at a glance

CriterionLangSmithLangfuseArize Phoenix
Overall rating6.88.07.2
Output qualityN/A (no evidence found)87
Ease of use787
Pricing value788
API and integration888
Problem fit586

All three scored 8 on integration, so the gap is elsewhere. LangSmith loses most of its ground on problem fit, where the only independent evidence we found (a comparison post from Future AGI) says it fits poorly outside LangChain stacks and takes real setup work. Phoenix loses points because there is little public evidence on onboarding, and because its GitHub repo shows 1,039 open issues against 11,654 stars at the time we scored it.

Feature comparison

FeatureLangSmithLangfuseArize Phoenix
Core tracingAgent tracing and observabilityHierarchical traces for LLM calls, tool calls and retrieval stepsTracing of agent runs, built on OpenTelemetry
EvaluationsYes, plus deployment toolingEvaluation metricsEvals, datasets and experiments
Prompt managementNot listed in our dataYesYes, with a playground
SDKsPython, TypeScript, Go, JavaSDKs plus OpenAI and LangChain integrationsOpenTelemetry instrumentation
OpenTelemetrySupportedNot listed in our dataNative base
Framework integrationsLangChain, LangGraph, OpenAI, Anthropic, LlamaIndexOpenAI, LangChainOpenAI Agents SDK, Claude Agent SDK, LangGraph, CrewAI, LlamaIndex, DSPy
FilteringNot detailed in our dataBy user, session, cost, latency, custom metadataNot detailed in our data
Self-hostingEnterprise plan only (hybrid or full VPC)Free, via Docker Compose, Kubernetes (Helm) or Terraform on AWS, GCP and AzureFree via pip, Docker or Helm; AX Enterprise also self-hostable
LicenceProprietary SaaSOpen sourceElastic License 2.0 (source-available)
Coding-agent accessNot listedNot listedRemote MCP server for Claude Code and Cursor
Data regionsUS, EU or APAC per our data; pricing page lists US or EU cloudCloud or your own infraLocal, SaaS or self-hosted
GitHub starsN/A (closed platform)About 35.2kAbout 11.9k

LangSmith

LangSmith has the widest SDK spread of the three, with Python, TypeScript, Go and Java, and it supports OpenTelemetry, so you are not locked to LangChain's own callbacks. Sign-up works with Google, GitHub, Discord or email. Where it fits best is LangChain and LangGraph. If your agent graph already runs on LangGraph, tracing is close to a config flag, and LangSmith's deployment tooling sits next to the traces. Outside that stack the case gets weaker. The Future AGI comparison we relied on reports significant setup overhead for non-LangChain code, and we found no independent evidence that contradicts it.

Langfuse

Langfuse gives you nested traces you can filter by user, session, cost, latency or any metadata you attach, along with prompt management and evals. Its self-hosting story is the most complete of the three. The project publishes guides for Docker Compose, Helm on Kubernetes and Terraform on all three major clouds, and the pricing FAQ says self-hosting is free. It has about 35.2k GitHub stars, and Langfuse says 21 of the Fortune 50 use it. One thing to watch is that Langfuse has joined ClickHouse. That gives it a well-funded parent, and it could also change the roadmap.

Arize Phoenix

Phoenix is the most agent-focused of the three on paper. It is built on OpenTelemetry and lists integrations with the OpenAI Agents SDK, the Claude Agent SDK, LangGraph, CrewAI, LlamaIndex and DSPy. You can start it locally with pip install arize-phoenix and phoenix serve, per the repo. Its remote MCP server stands out, because it lets Claude Code or Cursor query your traces and experiments directly. The licence matters here. The repo states Elastic License 2.0, which allows internal use and self-hosting but restricts offering Phoenix as a managed service to third parties. Most product teams will be fine with that, but have your legal team read it before you build on it.

Pricing comparison

PlanLangSmithLangfuseArize (AX)
FreeDeveloper: $0, 1 seat, up to 5k traces/monthHobby: $0, 50k units/month, 2 users, 30 days data accessAX Free: $0, 25k spans/month, 1 GB ingestion, 15 days retention
Entry paidPlus: $39 per seat/month, up to 10k traces/monthCore: $29/month, 100k units, unlimited users, 90 daysAX Pro: $50/month, 50k spans, 10 GB, 30 days retention
Mid tierNonePro: $199/month, 100k units, 3 years data accessStartup pricing on request
EnterpriseCustom, annual, adds hybrid or self-hosted$2,499/monthCustom, SaaS or self-hosted
Overage$2.50 per 1k traces, per a metaCTO pricing guide; LangChain's page lists pay-as-you-go without a rate$8 per 100k units, falling to $6 per 100k above 50MNot published
Free self-hostNoYesYes (Phoenix OSS)

The pricing models differ in shape. LangSmith charges per seat and then per trace. A five-person team on Plus pays $195 a month before overage, and per LangChain's pricing page, base traces keep 14 days of retention unless you pay to upgrade them to 180-day extended traces. Langfuse charges a flat platform fee with unlimited users and published, graduated overage. That makes it the easiest of the three to model in a spreadsheet. A ten-person team on Core pays $29 plus usage. The jump from Pro at $199 to Enterprise at $2,499 is steep. Pro also has a $300/month Teams add-on for SSO and fine-grained RBAC, which covers the common reason to upgrade.

Arize's AX Free tier is the smallest in volume at 25k spans and 15 days of retention, and AX Pro only doubles spans to 50k. The real cost lever is Phoenix OSS. If you can run a container, you pay for your own infra and nothing else. The units also differ. Langfuse counts observations (each span or event), LangSmith counts traces (whole runs) and Arize counts spans, so a raw comparison of 5k against 50k against 25k overstates how far apart the free tiers are.

Use case scenarios

You build on LangChain or LangGraph: pick LangSmith

First-party tracing for LangGraph is the main reason to pay LangSmith's per-seat price. You get eval and deployment tooling from the same vendor that ships your framework, plus SDKs in Go and Java if your services are polyglot. If you need data residency, LangChain offers US or EU cloud, and hybrid or full VPC hosting on Enterprise.

You want open source and self-hosting with predictable cost: pick Langfuse

Langfuse is the safest default for most teams. It is open source and free to self-host, and it has deployment guides for every major cloud and the largest community of the three by GitHub stars. Unlimited users on paid cloud plans means headcount does not drive the bill. The filters by user, session and cost help when you are chasing a cost spike to one customer. It works with LangChain too, so you keep the option to change frameworks later.

You build multi-framework agents or use AI coding tools: pick Arize Phoenix

If your stack mixes the OpenAI Agents SDK, CrewAI and DSPy, Phoenix's OpenTelemetry base and integration list cover more of it than the other two. The MCP server for Claude Code and Cursor is unique in this group and useful if your engineers debug from inside their editor. Phoenix also starts locally with one command, which suits notebook-driven eval work. Check that ELv2 suits your use, and plan for some trial time given the thin public evidence on onboarding.

You are a solo developer on a small budget

Langfuse Hobby gives the most free volume, at 50k units and 2 users. Phoenix running locally costs nothing and has no span cap. LangSmith's free tier is one seat and 5k traces, which is enough to try it on a LangGraph project and thin for anything else.

Verdict

Langfuse wins overall with an 8.0 rating. It has the broadest fit, the clearest pricing and the most mature self-hosting, and the main risks are the Enterprise price jump and whatever the ClickHouse move brings. Arize Phoenix comes second at 7.2 and is the pick for agent teams on mixed frameworks or those who want trace access from Claude Code and Cursor, provided the Elastic License works for them. LangSmith scores 6.8. It is the right choice when LangChain or LangGraph is already your foundation, and a harder sell elsewhere because per-seat billing and the reported setup overhead work against it. If you are unsure, start with Langfuse or Phoenix self-hosted, since neither costs anything to try on your own infra.

Frequently asked questions

Which is better, LangSmith, Langfuse or Arize Phoenix?

On the scored.tools score, Langfuse leads this comparison at 8/10, ahead of Arize Phoenix (7.2/10) and LangSmith (6.8/10). Scores come from the same five criteria for every tool and are not affected by payment.

How much does LangSmith cost?

LangSmith lists these plans: Free: Free; Plus: $39 per seat/month. Prices as listed on the vendor's site when we last updated this tool, 2026-09-29.

How much does Langfuse cost?

Langfuse lists these plans: Hobby: Free; Core: $29/month; Pro: $199/month. Prices as listed on the vendor's site when we last updated this tool, 2026-09-29.

How much does Arize Phoenix cost?

Arize Phoenix lists these plans: AX Free: $0/month; AX Pro: $50/month; Startup: Startup pricing available. Prices as listed on the vendor's site when we last updated this tool, 2026-09-29.

Scored this week

One email a week: new reviews, score changes and one pick. You confirm by email, and unsubscribe in one click.