Langfuse vs LangSmith in 2026: OTel, Evals & AI TCO
Engineering and 36-month TCO teardown of Langfuse vs LangSmith in 2026 alongside Arize Phoenix and Braintrust: ClickHouse OpenTelemetry ingestion, seat vs unit pricing, PII redaction, and MIT self-hosting.

On this page

Every engineering leader moving agent workflows from notebooks into customer production ends up at the Langfuse vs LangSmith question (also searched as LangSmith vs Langfuse, usually next to Arize Phoenix vs Langfuse and Braintrust). In a prototype, console.log feels like enough. In production, a multi-step agent can loop a dozen times over a malformed tool schema, run up a four-figure token bill overnight, and invent a refund policy while returning HTTP 200 OK.
APM tools such as Datadog, Sentry and New Relic check container health (see SigNoz vs Datadog for that side of the stack). They don't cover observability vs monitoring for non-deterministic AI agents and large language models, where failures hide in prompt context assembly, retrieval chunk relevance, tool-argument drift and judge rubrics. So teams end up picking among LLM observability tools, and a wrong pick costs you later: proprietary SDK lock-in, a $39 per seat fee that keeps PMs and domain reviewers out of the traces, and Enterprise-only self-hosting once InfoSec asks for HIPAA or sovereign VPC isolation.
At Idealogic, our AI software development, AI agent development, and generative AI development teams instrument production LLM pipelines in regulated fintech and healthtech environments. This comparison covers Langfuse, LangSmith, Arize Phoenix and Braintrust, the OpenTelemetry design we use for PII redaction and prompt fallback, and a 36-month TCO model for 28 seats and 25M agent spans a month.
The short version
- Langfuse is the pick for OpenTelemetry portability, free seats and MIT control inside your VPC. ClickHouse acquired it in January 2026 (Langfuse joins ClickHouse). Langfuse keeps its MIT-licensed core, ingests native OpenTelemetry GenAI semantic conventions, charges $0 per seat (unlimited users on $29 Core, $199 Pro and $2,499 Enterprise), and bills graduated units from $8 down to $6 per 100k.
- LangSmith gives you LangGraph state debugging with almost no setup, at a higher seat and retention cost. LangSmith renders LangGraph node transitions, thread state diffs and checkpoints out of the box. LangSmith pricing is $39 per seat per month on Plus (10k base traces included), then $5.00 per 1,000 (0.005 LSU) for 14-day base root traces and $7.50 per 1,000 (0.0075 LSU) for 180-day extended traces. Self-hosting in your VPC is an Enterprise feature.
- Arize Phoenix (Elastic License 2.0) is the stronger tool for OpenInference retrieval and embedding analysis, but it has no production prompt CMS workflow. Braintrust ($249 per month on Pro, no seat fee) is built around pull-request regression scoring and runs on a closed-source server.
- At 28 seats and 25M monthly agent spans with 90-day-plus retention under HIPAA/SOC 2, owned MIT Langfuse has the lowest modeled TCO. Over 36 months, self-hosted MIT Langfuse on AWS ECS and ClickHouse ($122,400, build and AWS included) comes in $41,100 below Langfuse Cloud Enterprise ($163,500), $96,900 below Braintrust Enterprise ($219,300) and $211,500 below LangSmith Enterprise ($333,900). The Enterprise figures are illustrative negotiated rates, not list prices, and Langfuse Cloud Pro undercuts all four if it covers your compliance scope.
LLM observability is tracing every prompt, retrieval, tool call and model response of an AI application, plus scoring the outputs, so teams can find why an agent failed. Langfuse and LangSmith are the two most-compared LLM observability tools: Langfuse is MIT-licensed and OpenTelemetry-first, LangSmith is a proprietary product built around LangChain and LangGraph.
Pricing verified 2026-11-14 against vendor pricing pages. They change often, so check again before you budget.
LLM observability tools compared: Langfuse vs LangSmith vs Arize Phoenix vs Braintrust
A Langfuse vs LangSmith evaluation comes down to four boundaries: ingestion protocol, storage engine, unit vs root-trace billing, and VPC deployment licensing.
1. Langfuse (a ClickHouse company): MIT core, ClickHouse OLAP and native OpenTelemetry
Langfuse is the better fit when your agents run on more than one SDK, when non-engineers need to read traces, or when the data has to stay in your own VPC.
Before joining ClickHouse in January 2026, Langfuse rewrote v3 around ClickHouse columnar storage, with PostgreSQL for projects, prompts and RBAC and S3 for blob offload. Multi-step agents emit large JSON prompts, tool schemas and RAG chunks, and row-oriented OLTP stores struggle with GROUP BY model, tenant_id queries over that data. Columnar storage with ZSTD compression suits repetitive prompt text and keeps P99 latency and token-spend aggregations fast.
What sets Langfuse apart in 2026:
- First-class OpenTelemetry (
@langfuse/oteland/api/public/otel). Through Langfuse's OpenTelemetry integration and@langfuse/otel, services that emitgen_ai.*spans from Vercel AI SDK telemetry, OpenAI, Anthropic, LangGraph, LlamaIndex, CrewAI or Go and Rust OTLP exporters land in one trace tree. You aren't tied to a vendor SDK. - Prompt management that's decoupled from deploys. Prompts are versioned with labels (
production,staging,canary) and cached in the SDK after the first fetch (default TTL 60 seconds), so the hot path doesn't wait on a network call. Each version links to trace scores, so you can compare hallucination rates between prompt v14 and v15. - MIT self-hosting with no license fee. Self-hosting Langfuse runs the same core tracing, prompt CMS, dataset, playground and LLM-as-a-Judge codebase under the MIT License. Only governance add-ons (SAML SSO, SCIM, audit logs, retention policies) need
LANGFUSE_EE_LICENSE_KEYor Cloud Pro/Enterprise.
2. LangSmith: LangGraph state inspection, LSU metering and the retention upgrade
LangSmith is the better fit when nearly everything you ship runs on LangChain and LangGraph. In a LangSmith vs Langfuse evaluation, its advantage is how tightly it's coupled to those two frameworks. Set LANGSMITH_TRACING=true and it captures cyclical graph transitions, multi-agent handoffs, node state diffs and thread memory checkpoints without extra code (a good complement to the durable orchestrators in Temporal vs Airflow). It also takes OpenTelemetry through an official OTLP endpoint.
Polyglot teams tend to hit friction in a few places, and that's usually when they start looking at LangSmith alternatives:
- Per-seat pricing. As agentic AI in production shows, raising agent accuracy needs clinicians, compliance analysts, support leads and PMs grading real failure trajectories every week. At $39 per seat per month, 28 collaborators cost $13,104 a year before you log one trace.
- Root-trace LSU metering. LangSmith bills per root trace (1 LSU = $1.00). Plus includes 10,000 shared base traces, then charges 0.005 LSU ($5.00 per 1,000) for 14-day base traces and another 0.0025 LSU ($2.50 per 1,000, so $7.50 per 1,000 in total) for 180-day extended retention.
- VPC self-hosting sits on Enterprise. Running LangSmith inside your own Kubernetes cluster for HIPAA BAA coverage or EU banking sovereignty means an annual LangSmith Enterprise contract with custom pricing. The pricing page doesn't state BAA terms, so confirm them with LangChain.
Watch the retention auto-upgrade. Per the LangSmith data retention docs, a base trace moves to extended retention when feedback is sent through the API or SDK with
extend_trace_retention=true, or when an online evaluator or automation rule with retention extension enabled touches it. UI feedback doesn't trigger it, and you can opt evaluators out. Unchecked, $5.00 per 1,000 logs become $7.50 per 1,000 invoices.
LangSmith pricing vs Langfuse pricing in 2026
LangSmith pricing is seat-based with per-trace overages. Langfuse pricing is a flat plan fee with per-unit overages and no seat charge.
Langfuse Hobby has hard limits. The free Hobby plan covers 2 users and 50k units a month with 30-day data access. Fine for a prototype, too small for a team review workflow.
3. Arize Phoenix vs Langfuse and Braintrust: where the specialists fit
If you're shopping for Langfuse alternatives, these are the two worth a serious look.
- Arize Phoenix vs Langfuse. Arize Phoenix (Elastic License 2.0, source-available) runs natively on OpenTelemetry through OpenInference and is strongest at embedding and retrieval analysis for drift across Pinecone vs Qdrant RAG corpora. Langfuse is the better choice when you need a production prompt CMS, permissive MIT licensing, annotation queues and ClickHouse-scale cost rollups without moving to commercial Arize AX.
- Braintrust vs Langfuse. Braintrust is free up to 1 GB, then $249 per month on Pro with unlimited users, metered on processed GBs and scores. Its strength is pull-request eval comparisons in CI (see the Braintrust docs). Pick Braintrust when hosted PR experimentation matters more than open-source control, and Langfuse when regulated workloads need MIT self-hosting inside your VPC. Braintrust vs LangSmith is a simpler split: CI evals against LangGraph debugging.
Langfuse vs LangSmith in production: OpenTelemetry PII redaction and prompt fallback
Shipping raw LLM telemetry from a healthtech or fintech agent straight to an external SaaS collector widens your HIPAA minimum-necessary exposure (45 CFR 164.502(b) and 164.514(d)), and without a signed BAA it's an impermissible disclosure. It also sits badly with GDPR Article 25, data protection by design, because user prompts routinely carry Social Security numbers, IBANs, JWTs and clinical notes. There's a second risk. If you fetch managed prompts over HTTP with no local fallback, an observability blip becomes an HTTP 500 outage.
This is the setup we use on LLM architecture and AI agents for business engagements. The decisions that matter most:
The core of it is the mask hook, a function that runs in-process and replaces emails, SSNs and Bearer tokens with redaction markers before any span leaves the service. A string-only version is a few lines of regex. In production we also walk nested objects and redact by key name, and you should check that the mask covers the AI SDK's gen_ai.* attributes (see Langfuse's TypeScript SDK advanced usage).
The processor reads its keys and base URL from the standard Langfuse environment variables. For the agent step itself, Vercel AI SDK streamText emits standard gen_ai.* child spans, and its onFinish callback is where token cost lands on the parent span. One version note: experimental_telemetry is the AI SDK v5/v6 syntax. Newer releases use telemetry with @ai-sdk/otel, as the AI SDK telemetry docs show. Take model IDs and token rates from Anthropic's models and pricing pages rather than hardcoding them.
Langfuse vs LangSmith 36-month TCO: 28 seats and 25M spans per month
Here's the unit math against the root-trace math for a Series B agent fleet over 36 months, with 90-day-plus retention under HIPAA/SOC 2:
- Team: 28 active seats (12 AI/backend engineers, 6 PMs, 6 clinical or financial SMEs, 4 QA engineers).
- Volume: 1,250,000 root traces a month. Each averages 18 child spans (router, RAG, reranker, tools, guardrails, LLM steps) plus 1 root trace and 1 judge score, which makes 20 Langfuse units per trace and 25,000,000 units a month.
- Governance: 90-day to 180-day-plus retention, SAML SSO, SOC 2 Type II, and a HIPAA BAA or EU (
eu-central-1) residency.
Engineering labor uses Idealogic's own rates: $75/hr senior CEE against $165/hr for a US agency. For context, EU whole-economy labour cost averaged EUR 34.9/h in 2025 (Eurostat lc_lci_lev), and O*NET / BLS (15-1252.00) tracks US software developer wages. Rates match AI agent development cost.
Read the Enterprise rows as a model. All three vendors quote Enterprise on request. The $8,600, $5,400 and $3,850 monthly figures are illustrative negotiated rates, not list prices (the Langfuse one assumes about 8% off its $4,176 list math).
- Option A: LangSmith Enterprise (28 seats, 1.25M extended root traces a month, HIPAA BAA)
- Setup (72 hrs at $75/hr = $5,400): LangGraph/OTel setup, SSO/RBAC, PII proxy and golden datasets.
- 36-month Enterprise contract ($8,600/mo = $309,600): At Plus list rates, seats are 28 x $39 = $1,092/mo and 1,240k extended 180-day traces (1.25M less the 10k included, at 0.0075 LSU or $7.50 per 1,000) are $9,300/mo, so $10,392/mo at list without BAA or SSO. We model an annual LangSmith Enterprise contract with BAA, SSO/ABAC and volume LSUs at $8,600/mo ($103,200/yr), plus 7 hrs/mo of ops ($18,900) = $333,900.
- Option B: Braintrust Enterprise (28 unlimited seats, 25M spans and scores, hybrid data plane with HIPAA BAA)
- Setup (80 hrs at $75/hr = $6,000): SDK/OTel converters, GitHub PR eval gates and hybrid data plane setup.
- 36-month Enterprise contract ($5,400/mo = $194,400): No seat fee. The modeled contract covers processed JSON trace GBs, 1.25M judge scores, SAML SSO and the HIPAA hybrid data plane ($64,800/yr), plus 7 hrs/mo of ops ($18,900) = $219,300.
- Option C: Langfuse Cloud Enterprise (no seat fee for 28 users, 25M monthly units, HIPAA US or EU region)
- Setup (80 hrs at $75/hr = $6,000):
@langfuse/otelLangfuseSpanProcessor, PIImaskhook, prompt fallback cache and SAML SSO. - 36-month Cloud Enterprise ($3,850/mo = $138,600): The 28 users cost $0. List math is the Enterprise base ($2,499/mo with 100k units, 3-year access, SCIM and audit logs) plus graduated units on 25M (900k at $8 per 100k = $72, 9M at $7 per 100k = $630, 15M at $6.50 per 100k = $975, so $1,677/mo), which totals $4,176/mo at list. We model $3,850/mo on an annual commit, plus 7 hrs/mo of ops ($18,900) = $163,500.
- The Pro variant. Enterprise is assumed here for SCIM, audit logs and enterprise SSO, not for the BAA. Langfuse's HIPAA page lists the HIPAA region and a signed BAA for Pro or higher, though its pricing matrix shows that region under Enterprise, so confirm eligibility. If Pro is enough, Cloud Pro ($199) plus Teams ($300) costs $2,176/mo, or $103,236 over 36 months.
- Setup (80 hrs at $75/hr = $6,000):
- Option D: Owned self-hosted MIT Langfuse v3 on AWS ECS, ClickHouse and S3 with a CI eval pipeline (Idealogic)
- Turnkey build (720 hrs at $75/hr CEE = $54,000, or $118,800 at $165/hr US): HA
langfuse-webandlangfuse-workeron HIPAA-eligible AWS services (ECS Fargate, RDS, S3, KMS), replicated ClickHouse with ZSTD, Multi-AZ RDS Postgres 17, Valkey, S3 archive, OTel PII redaction, GitHub CI evals and runtime budget guardrails. No license, seat or span fees. - 36-month AWS VPC infra ($1,150/mo = $41,400): Two
r7g.xlargeClickHouse nodes plus EBS (about $540/mo, indicative), ECS and the OTel Collector ($240/mo), Multi-AZ RDS and Valkey ($220/mo), and S3 3-year tiering with ALB/KMS ($150/mo), all under your AWS HIPAA BAA, plus 10 hrs/mo of ops ($27,000) = $122,400. The AWS lines are estimates, so price your own region.
- Turnkey build (720 hrs at $75/hr CEE = $54,000, or $118,800 at $165/hr US): HA
Crossover rule: when Langfuse Cloud Pro beats self-hosting
Self-hosting isn't always the answer. If Langfuse Cloud Pro covers your compliance scope, Pro ($199/mo) plus Teams ($300/mo) costs $2,176/mo at 25M monthly units, a 36-month all-in TCO of $103,236. That beats self-hosted AWS TCO ($122,400) until volume passes roughly 33M to 38M monthly units.
Self-hosting MIT Langfuse wins right away in two cases: (1) HIPAA, PCI DSS v4.x or EU banking rules forbid prompt egress outside your VPC, or (2) the agent fleet passes about 35M spans a month.
2026 AI observability, CI eval and guardrail build costs (senior CEE vs US)
Our AI agent development and AI software development teams ship fixed-milestone observability stacks at $75/hr senior CEE against $165/hr US agency rates:
Inside the worked 720-hour ($54,000 CEE vs $118,800 US) self-hosted MIT Langfuse and CI eval build
- 1. Multi-AZ AWS ECS, ClickHouse Keeper and Terraform IaC (30% / 216 hrs / $16,200 CEE vs $35,640 US): Private-subnet ClickHouse with ZSTD, Multi-AZ RDS Postgres 17, Valkey, S3 tiering, KMS and automated snapshot restores under your AWS BAA.
- 2. OpenTelemetry
gen_ai.*pipeline and PII/PHI redaction (25% / 180 hrs / $13,500 CEE vs $29,700 US): Non-blocking@langfuse/oteland Python OTLP exporters, regex plus NER entity scrubbing before export, and tenant correlation IDs. - 3. CI regression eval harness and position-swapped LLM judges (20% / 144 hrs / $10,800 CEE vs $23,760 US): 50+ golden trajectories, tool-schema assertions, calibrated dual-order LLM judges and
pass^kPR gates. - 4. Versioned prompt CMS fallback and runtime budget guardrails (15% / 108 hrs / $8,100 CEE vs $17,820 US): SDK-cached prompt resolution with Git fallbacks, per-tenant USD token caps and loop circuit breakers (paired with hot-path routing in Portkey vs LiteLLM).
- 5. FinOps token dashboards, drift alerts and k6 load tests (10% / 72 hrs / $5,400 CEE vs $11,880 US): Cost-per-task alerts, Slack/PagerDuty latency triggers and span ingestion load tests.
Never wire proprietary tracing decorators across fifty agent tools. Emit OpenTelemetry GenAI spans, scrub PII in-process before export, and back every managed prompt with a Git fallback. You can ship on Langfuse Cloud Core for twenty-nine dollars a month in Sprint 1, and lift the identical MIT ClickHouse stack into an air-gapped HIPAA VPC when enterprise procurement arrives.
Shipped Idealogic proof and 5-step CTO checklist
Idealogic ships observable, cost-bounded AI and regulated products:
- LLM + pgvector SaaS (Pressmerit): Next.js + NestJS + PostgreSQL + pgvector ranking and outreach across 26 industries with tenant cost tracking.
- ISO 27001 audit trails (Chaindoc): Shipped in 22 weeks across 500+ organizations and 10,000+ signed agreements at 99.9% uptime.
- HIPAA and GDPR isolation (AddMed): Zero-PHI leak telemetry and strict healthcare RBAC with zero compliance incidents.
Five checks settle your Langfuse vs LangSmith decision for 2026:
- All-in on LangGraph, fewer than 8 engineers, 14-day debugging? Pick LangSmith Plus ($39 per seat per month).
- Polyglot SDKs (Vercel AI SDK, OpenAI, Anthropic, CrewAI) with 15 to 40+ collaborators and no PHI? Pick Langfuse Cloud Core or Pro. You pay no seat fee and get OTel ingestion.
- Local RAG embedding drift analysis under ELv2, or hosted PR CI evals? Pick Arize Phoenix or Braintrust.
- HIPAA PHI that can't leave your VPC, sovereign banking data, or 35M+ monthly spans? Go Langfuse self hosted on AWS ECS and ClickHouse (about $1,150/mo of AWS under your BAA). If a managed HIPAA region is acceptable, price Cloud Pro first.
- Want a 2-week audit of agent token burn and evals? Contact our AI engineering team for a scoped Blueprint.

CTO and Co-founder at Idealogic. Directing engineering, systems architecture, security, and full-stack delivery across web, mobile, and blockchain.
Frequently asked questions
Choose Langfuse when your stack mixes SDKs such as the Vercel AI SDK, OpenAI, Anthropic, LlamaIndex or CrewAI, when you want vendor-neutral OpenTelemetry GenAI instrumentation, when product and domain reviewers need access without per-seat fees, or when you need MIT-licensed self-hosting inside your own AWS or GCP VPC. Choose LangSmith when your team builds mostly on LangGraph and LangChain and wants state graph debugging, checkpoint replay and hosted Agent Server integration with almost no setup.
ClickHouse acquired Langfuse in January 2026. Langfuse v3 already ran on ClickHouse columnar storage for LLM spans and token aggregations, so the fit was technical. After the deal, the Langfuse maintainers confirmed that core tracing, prompt management, datasets and evaluations stay MIT-licensed and free to self-host on Docker or Kubernetes without event caps.
LangSmith pricing on Plus is $39 per seat per month with 10,000 shared base traces included. Overages are metered in LangChain Standard Units at $1 per LSU: 0.005 LSU ($5.00 per 1,000) for 14-day base root traces, plus 0.0025 LSU ($2.50 per 1,000) to upgrade a trace to 180-day extended retention, so an extended trace costs $7.50 per 1,000. Langfuse pricing has no seat fee. Core is $29 per month with 90-day data access, Pro is $199 with 3-year access and Enterprise is $2,499, all with unlimited users. Langfuse meters billable units (traces, child observations and scores) from $8 down to $6 per 100,000 units.
LangSmith bills per root trace, however many tool calls or retrieval spans run under it. A base trace moves from 14-day to 180-day extended pricing ($7.50 per 1,000 traces) when feedback is sent through the API or SDK with extend_trace_retention=true, or when an online evaluator or automation rule with retention extension enabled touches it. You can opt evaluators out. Langfuse bills every root trace, child observation and score as its own unit, but at $6.50 to $8 per 100,000 units a 20-unit agent trace costs roughly $0.0013 to $0.0016, against $0.0075 for an extended LangSmith trace.
Arize Phoenix is an OpenTelemetry and OpenInference workbench for RAG embedding drift analysis, local notebook experiments and self-hosted tracing. It ships under the source-available Elastic License 2.0, which isn't an OSI-approved license. Langfuse has an MIT core and adds collaborative prompt management, SDK-side prompt caching, human annotation queues and production token-cost dashboards on top of ClickHouse tracing.
Braintrust fits teams whose workflow centers on pre-merge CI regression gates, fast dataset experiments and score comparisons across pull requests. Its Pro plan costs $249 per month with unlimited seats and is metered on processed gigabytes and scores. The Braintrust server is closed-source SaaS, with a hybrid data plane on Enterprise, while Langfuse gives regulated teams the full MIT source to run inside an air-gapped VPC.
LangSmith offers self-hosted Kubernetes deployment, hybrid VPC options and custom SSO on Enterprise contracts with custom pricing. A HIPAA Business Associate Agreement is handled through Enterprise sales, so confirm it with LangChain. Langfuse self hosted works differently: regulated healthcare and financial teams can run the MIT-licensed web, worker, PostgreSQL and ClickHouse stack inside an existing AWS HIPAA BAA or a Frankfurt GDPR VPC with no software license fee. On Langfuse Cloud, the dedicated HIPAA region with a signed BAA is listed for the Pro plan or higher, so confirm plan eligibility with Langfuse before you sign.
At Idealogic's senior Central and Eastern European rate of $75 per hour (against $165 per hour for a US agency), a 2 to 3 week OpenTelemetry instrumentation and eval audit sprint costs $13,500 to $21,000. A Langfuse or LangSmith Cloud plus CI eval integration runs $24,000 to $39,000. A turnkey 720-hour self-hosted MIT Langfuse cluster on AWS ECS and ClickHouse with PII redaction and CI gates costs $54,000, or $118,800 at US rates.
More from the journal

Algolia vs Typesense vs Meilisearch: 2026 Search & TCO
Algolia vs Typesense vs Meilisearch in 2026, compared on engineering and 36-month TCO: Algolia pricing per 1,000 autocomplete requests vs in-RAM Raft clusters, Rust LMDB disk trade-offs, and blue/green reindexing.

Portkey vs LiteLLM in 2026: AI Gateways, Latency & TCO
Engineering and 36-month TCO comparison of Portkey vs LiteLLM vs Helicone and compiled Go/Rust AI gateways in 2026: streaming overhead, enterprise SSO walls, PostgreSQL 17 FORCE RLS token-budget ledgers, and 120M req/mo math.

Pinecone vs Qdrant vs Weaviate: 2026 RAG & TCO Guide
Engineering and 36-month TCO benchmark of Pinecone vs Qdrant vs Weaviate and PostgreSQL pgvector in 2026: filterable HNSW recall, hybrid BM25/RRF search, multi-tenant isolation, HIPAA BYOC, and RAG build costs.