Arize Phoenix vs Langfuse vs LangSmith 2026: Which to Use
The short answer
Langfuse for the self-hosted default, Arize Phoenix for OpenTelemetry-native evaluation (and now for Dynatrace shops), LangSmith for LangGraph depth. The three are converging on the same feature list — traces, prompts, datasets, evals, agent views — so the decision comes down to three questions: can it run on your infrastructure, does it bias your framework choice, and who owns it now. As of October 2026 the ownership answers are Dynatrace, ClickHouse and LangChain respectively. Facts verified October 2, 2026.
Side by side
| Arize Phoenix | Langfuse | LangSmith | |
|---|---|---|---|
| Owner (Oct 2026) | Dynatrace (acquisition closed Oct 1, 2026, ~$915M) | ClickHouse (since January 2026) | LangChain Inc. |
| Licence | Open source (Elastic/ELv2) | Open source (MIT core) | Proprietary |
| Self-host | Yes — container or pip install | Yes — Docker, Postgres + ClickHouse | Enterprise only |
| Hosted tier | AX Free: 25K spans/mo, 1 GB, 15-day retention, $0 | Hobby $0 (50K units) · Core $29 · Pro $199 · Enterprise $2,499 /mo | Developer $0 (1 seat, 5K traces) · Plus $39/seat/mo (10K traces) · Enterprise custom |
| Billing unit | Spans | Units (≈ one observation) | Seats + base traces + usage units |
| Instrumentation | OpenTelemetry / OpenInference first | SDKs + OTel ingestion | LangChain/LangGraph native; SDK for others |
| Framework bias | None | None | LangGraph |
| Evals | Strong: LLM-as-judge library, experiments, datasets | Datasets, experiments, LLM/human/custom scores | Datasets, experiments, online evals |
| Prompt management | Yes | Yes — versioning is a core feature | Yes (Prompt Hub) |
| Agent graph view | Trace tree | Trace tree, sessions | Node-level LangGraph state transitions |
| Scale signal | 4,000+ enterprises, ~1T spans/month (Phoenix) | Backed by ClickHouse’s analytics engine | Default for the LangChain ecosystem |
| Best for | OTel purity, eval rigour, Dynatrace customers | Self-hosted default, prompt versioning | LangGraph debugging |
Prices read from vendor pages on October 2, 2026; usage overages apply on every hosted tier.
Arize Phoenix — OpenTelemetry-native, now part of Dynatrace
Phoenix is the open-source half of Arize (the commercial platform is Arize AX). Its design decision was to build on OpenTelemetry with the OpenInference semantic conventions, which means traces you emit for Phoenix are portable to any OTel backend — the lowest lock-in of the three. Its evaluation library is the most opinionated: prebuilt LLM-as-judge evaluators for hallucination, relevance, toxicity and tool-call correctness, plus datasets and experiments for regression testing.
The news is ownership. Dynatrace closed its acquisition of Arize on October 1, 2026 for roughly $915M ($815M cash plus replacement equity). Arize reports Phoenix in use at 4,000+ enterprises processing about a trillion spans per month. Dynatrace wants one platform that evaluates agents in development and monitors them in production, correlated with the APM, logs and infrastructure data it already holds. For a Dynatrace customer this is the obvious choice. For everyone else, the question is roadmap priority: Phoenix stays open source, but expect the best integrations to land Dynatrace-side first. Arize’s AX Free hosted tier (25K spans, 15-day retention, Signal limited to 10 issues/month) is enough to evaluate the product; production volumes move to paid AX or self-hosted Phoenix.
Langfuse — the self-hosted default
Langfuse covers the whole loop — traces, observations, sessions, prompt versioning, cost tracking, datasets, experiments, LLM/human/custom evaluation scores — and runs on your infrastructure at no licence cost. Since January 2026 it has been part of ClickHouse, which matters technically: Langfuse’s trace store is ClickHouse, so the acquisition aligns the product with its own database rather than redirecting it, and it should age well at high trace volume.
Langfuse Cloud bills per unit (roughly one span, generation or event): Hobby $0 with 50,000 units a month, 30-day data access and 2 users; Core $29/month and Pro $199/month, each including 100,000 units with $8 per additional 100,000; Enterprise $2,499/month. The difference between Core and Pro is retention, users and features, not volume. Self-hosting is free in licence terms only: you run Postgres and ClickHouse, own storage growth and handle upgrades. Prompt versioning is the underrated reason to pick it — prompts are code that lives outside your repository, and without versioning a quality regression is nearly impossible to attribute. Full ranking context in best LLM observability tools 2026.
LangSmith — deepest if you live in LangGraph
LangSmith is built by the LangChain team, and for LangGraph agents nothing else shows node-level graph execution, state transitions and branch decisions as clearly. It works with non-LangChain code via its SDK and accepts OpenTelemetry, so it is not strictly locked in — but the reason to choose it is the depth you only get inside that ecosystem, plus its bundled deployment tier for LangGraph agents.
Pricing is seat-based: Developer $0 for one seat and 5,000 base traces; Plus $39 per seat per month with 10,000 base traces and one dev-sized deployment, then usage billing in two units (traces and compute); Enterprise custom. Three engineers on Plus is $117/month before traces — fine for a small team, a real line item at 30 seats. LangSmith is cloud-only below Enterprise; if traces cannot leave your network, it is not an option without a negotiated deployment. For eval-first alternatives compared against it, see LangSmith vs Weights & Biases vs DeepEval.
Lock-in, honestly
| Risk | Phoenix | Langfuse | LangSmith |
|---|---|---|---|
| Data lock-in | Lowest — OTel/OpenInference traces are portable | Low — open schema, self-hosted DB | Medium — export exists, but graph views are proprietary |
| Framework lock-in | None | None | Value concentrates in LangGraph |
| Vendor-strategy risk | Dynatrace integration may outrun standalone roadmap | ClickHouse alignment is additive | Tied to LangChain’s commercial trajectory |
| Exit cost | Point the OTel exporter elsewhere | Export from your own DB | Re-instrument |
How to choose
- Traces contain customer data that cannot leave your network: Langfuse or Phoenix, self-hosted. LangSmith only via Enterprise.
- You run Dynatrace for APM: Arize Phoenix/AX — one pane, and the integration is now the vendor’s strategy.
- LangGraph is your orchestration layer and agent debugging is your daily pain: LangSmith.
- You want OpenTelemetry portability above all: Phoenix.
- You want one tool that does tracing, prompts and evals well enough, with the least ceremony: Langfuse.
- Budget-first solo developer: all three have a $0 tier; Langfuse Hobby’s 50,000 units is the most generous for tracing volume, LangSmith Developer the most generous for a LangGraph project.
For the broader landscape — Braintrust’s eval-first approach, Helicone’s proxy model, Datadog’s incumbent play — see Langfuse vs Braintrust vs Helicone and guardrails vs evals vs observability.
Last verified: October 2, 2026. Pricing from arize.com, langfuse.com and langchain.com; acquisition terms from Dynatrace’s October 1, 2026 announcement.