Langfuse vs LangSmith vs Braintrust (and PostHog): LLM Observability Pricing 2026
The six LLM observability tools buyers compare most, priced from each vendor's own pricing page, with a worked monthly bill at 300,000 traces and a five-person team.
The short answer
Choose Langfuse if you want an MIT-licensed core you can self-host, or a $29/month cloud plan billed per unit. Choose LangSmith if your agents run on LangChain or LangGraph and the team is small: it bills $39 per seat plus 0.005 LSU ($0.005) per base trace after 10,000. Choose Braintrust ($249/month Pro) if evaluation is the main job. Choose PostHog if LLM cost has to sit next to product analytics.
LLM observability is not product analytics
Product analytics tools such as PostHog and Mixpanel count what users do: page views, funnels, retention. LLM observability tools record what your model and agent did on each request. That means the prompt, every tool call and retrieval step as a nested span, token counts, cost, latency, and a score for whether the answer was any good.
People search "Langfuse vs PostHog" because the two now overlap. PostHog sells an AI Observability product that captures "generations, spans, and traces" as ordinary PostHog events and links each trace to a session recording. Langfuse goes the other way, with export integrations into PostHog and Mixpanel. The useful question is not which category you need. It's where your engineers will debug a bad answer, and where finance will look for token spend per customer.
The three dedicated tools also differ in what they bill. Langfuse bills data points, LangSmith bills people and traces, and Braintrust bills gigabytes and scores. That difference drives the cost gap in the worked example below.
Pricing at a glance
Prices and plan names are as published on each vendor's own pricing page; last verified 10 November 2026.
| Vendor | Billing unit | Published entry price | What it's for |
|---|---|---|---|
| Langfuse | "Units": every trace, observation (span, event, generation) and score | Hobby free (50k units, 2 users); Core $29/month with 100k units, then $8 per 100k | Open-source tracing, prompt management, evals |
| LangSmith | Seats plus base traces; usage metered in LSUs (1 LSU = $1) | Developer $0 (1 seat, 5k base traces); Plus $39 per seat/month with 10k base traces | Tracing and evals, tightest fit with LangChain and LangGraph |
| Braintrust | Processed data (GB) plus scores | Starter $0 (1 GB, 10k scores); Pro $249/month (5 GB, 50k scores) | Eval-first: experiments, scorers, release gates |
| PostHog AI Observability | Events | First 100,000 events free each month, then from $0.000350 per event | LLM traces joined to product analytics and session replay |
| Comet Opik | Spans | Free cloud (25k spans); Pro $19/month (100k spans, then $5 per 100k) | Apache-2.0 open-source tracing and evals |
| Arize AX | Spans plus ingestion GB | AX Free (25k spans, 1 GB); AX Pro $50/month (50k spans, 10 GB) | Managed platform; Phoenix is the self-hosted sibling |
Langfuse
Langfuse's cloud plans are Hobby (free), Core ($29/month), Pro ($199/month) and Enterprise ($2,499/month). Every paid plan includes 100,000 units. After that, the per-100k rate steps down: $8 from 100k to 1M units, $7 from 1M to 10M, $6.50 from 10M to 50M, and $6 above 50M. Core and above include unlimited users. What changes between plans is how far back you can see data: 30 days on Hobby, 90 days on Core, 3 years on Pro. On Langfuse Cloud, enterprise SSO, SSO enforcement and fine-grained RBAC need the $300/month Teams add-on on top of Pro.
Langfuse's self-hosted pricing page says you can "self-host all core Langfuse features for free without any limitations" under the MIT license. The repository is more exact. Code outside the `ee` directories is MIT. Enterprise features such as project-level RBAC, data retention policies and audit logs are paid. Ownership: ClickHouse acquired Langfuse. Langfuse announced it on 16 January 2026 and said "there are no planned changes to licensing." The repository copyright now reads ClickHouse, Inc. Self-hosted Enterprise (audit logs, project-level RBAC, retention policies) is now sold bundled with a ClickHouse Cloud, BYOC or Private plan and priced on top of it. The free self-hosted edition includes enterprise SSO.
When not to use Langfuse: when you have no one to run it and your volume is high. A span-heavy agent produces many units per user request, because each observation and score counts. Self-hosting removes the unit bill, but you take on running ClickHouse and the rest of the stack yourself.
LangSmith
LangSmith's Developer plan is free for one seat and 5k base traces a month. Plus costs $39 per seat per month, includes 10k base traces, and allows unlimited seats. Enterprise is custom. Extra traces cost "0.005 LSU/base trace" and "0.0025 LSU/extended trace upgrade" (the page sets 1 LSU = $1). That works out to $5 per 1,000 base traces. Base traces are kept 14 days and extended traces 180 days. A seat is "a distinct user inside your organization", and invited users count.
A trace is billed once however many steps it contains, which is the strongest part of LangSmith's model for deep agents. LangChain is not required. The quickstart traces plain OpenAI, Anthropic or Gemini clients with a wrapper or the `@traceable` decorator. Self-hosted and hybrid deployment are Enterprise-only.
When not to use LangSmith: when many people need access (every invited user is a $39 seat), or when you need traces kept past 14 days at volume, since each extended upgrade adds cost. It's also the wrong pick if you need self-hosting without an enterprise contract.
Get one new AI workflow every week
Practical playbooks, prompts, and tool stacks — for the SMB operators and enterprise teams actually shipping AI.
Braintrust
Braintrust bills processed data and scores, not traces or seats. Starter is $0 with 1 GB of processed data (then $4/GB), 10K scores (then $2.50 per 1,000) and 14-day retention. Pro is $249/month with 5 GB (then $3/GB), 50K scores (then $1.50 per 1,000) and 30-day retention, extendable at $0.50 per GB per month. Users are unlimited on every plan. Processed data is measured at ingest, so deleting data after you pass the quota doesn't lower the bill. On-prem deployment and a BAA are Enterprise-only.
When not to use Braintrust: when you want to score every production trace. Scores are the line that grows fastest, so sample. It's also a poor fit if you need HIPAA coverage below Enterprise.
PostHog, Opik and Arize, briefly
PostHog charges AI Observability events on a sliding scale: 100k free, $0.000350 per event up to 150k, falling in steps to $0.000060 per event above 50 million. There are no per-seat charges. Use it when the question is "which customers drive our token spend, and did the AI feature change retention?" See Langfuse vs PostHog and LangSmith vs PostHog.
Comet Opik is Apache-2.0 and self-hostable from the same codebase as its cloud. Its cloud Pro plan is $19/month for up to 50 team members and 100k spans, then $5 per 100k spans, with 60-day retention.
Arize AX Pro is $50/month for 50k spans, 10 GB and 30-day retention. The pricing page publishes no per-span overage rate. Phoenix, Arize's open-source option, uses the Elastic License 2.0. That license lets you self-host, modify and redistribute it, but forbids offering Phoenix to third parties as a hosted or managed service.
Worked example: 300,000 traces a month, five engineers
Assumptions: an agent handles 300,000 requests a month. Each request is one trace with five child observations (LLM generations, a retrieval and tool calls). An LLM-as-judge scores a 10% sample (30,000 scores). Processed data averages 10 KB per trace, or 3 GB a month. Five engineers need access. The judge runs inside each tool where the tool supports it, so on LangSmith its 30,000 evaluator runs are billed traces (a trace can be "an agent, evaluator, or playground session"), and on Langfuse each judge execution stores its own trace and generation. Model tokens for the judge are excluded everywhere.
| Tool and plan | Arithmetic | Monthly total |
|---|---|---|
| Langfuse Core | Units = 300k + 1.5M + 30k scores + 60k judge-execution units (1 trace + 1 generation each) = 1.89M. Overage: 900k × $8/100k = $72; 890k × $7/100k = $62.30. $29 + $134.30 | $163.30 |
| Langfuse Pro | $199 + $134.30 (same unit math); 3-year data access | $333.30 |
| LangSmith Plus | 5 seats × $39 = $195; (300k + 30k evaluator runs − 10k) × $0.005 = $1,600; 30k scored traces auto-upgraded to extended retention (on by default for new online evaluators) × $0.0025 = $75 | $1,870 |
| Braintrust Pro | 3 GB is under the 5 GB included; 30k scores are under the 50k included | $249 |
| Braintrust Starter | (3 − 1) GB × $4 = $8; (30k − 10k) × $2.50/1k = $50; 14-day retention | $58 |
| PostHog | 6 events per trace (1 trace + 5 spans/generations; eval events excluded, as PostHog's pricing page doesn't say whether eval results are billed events) = 1.8M events. $17.50 + $17.50 + $25 + $40 + (800k × $0.00007 = $56) | $156 |
| Comet Opik Pro | 1.5M spans − 100k included = 1.4M × $5/100k = $70; $19 + $70 | $89 |
Three inputs move these numbers most. On LangSmith, online evaluators upgrade the traces they score to 180-day retention by default (you can opt out per evaluator), which is the $75 in the table. Opting out brings it to $1,795. A judge run outside LangSmith that only posts feedback brings it to $1,645. On Braintrust, scoring every trace instead of 10% means 300,000 scores, or (300k − 50k) × $1.50/1k = $375 on top of Pro. Check your own span count per trace before comparing. A trace with 20 spans instead of 5 roughly triples the Langfuse, PostHog and Opik lines (to about $478, $471 and $314) but leaves LangSmith unchanged. Braintrust bills data ingested, and 10 KB per trace is this example's assumption, not a vendor figure. At 50 KB per trace (15 GB), Starter rises to (14 × $4) + $50 = $106 and Pro to $249 + (10 × $3) = $279.
How to choose
- Small team, LangGraph in production, modest volume: LangSmith. Per-trace billing ignores agent depth.
- Open source or data residency required, or high span counts: Langfuse. Self-host the MIT core, or use Core/Pro if you'd rather not run ClickHouse. See Langfuse vs LangSmith.
- Releases gated on eval scores: Braintrust, with sampled scoring. Pair it with the agent evaluation regression harness workflow.
- Product team that already runs PostHog: start with PostHog AI Observability and add a dedicated tool only when engineers outgrow it.
- Enterprise buyers: confirm SSO, audit logs and BAA plan placement first. On Langfuse Cloud, SSO is a $300/month Teams add-on to Pro (the free self-hosted edition includes it), Braintrust's BAA is Enterprise-only, and LangSmith's self-hosting is Enterprise-only.
Frequently asked questions
Is Langfuse free to self-host?
Yes. Langfuse's self-hosted pricing page offers the core platform free under the MIT license, with unlimited usage. Enterprise features such as audit logs, project-level RBAC and data retention policies are paid.
Is LangSmith only for LangChain?
No. LangSmith's quickstart traces plain OpenAI, Anthropic or Gemini clients with a wrapper or the `@traceable` decorator. LangChain and LangGraph just need fewer setup steps, a single environment variable.
How much does LangSmith cost per trace?
The LangSmith pricing page lists 0.005 LSU per base trace after the included allowance, with 1 LSU equal to $1, so $5 per 1,000 base traces. Upgrading a trace to 180-day extended retention adds 0.0025 LSU.
Who owns Langfuse now?
ClickHouse. Langfuse announced the acquisition on 16 January 2026 and said there were no planned licensing changes and that Langfuse Cloud keeps running.
Can PostHog replace Langfuse or LangSmith?
For cost tracking, trace inspection and LLM-as-judge evals tied to user behaviour, often yes. PostHog's first 100,000 events a month are free. Teams that need prompt versioning, dataset experiments or a dedicated eval workflow may still add Langfuse, LangSmith or Braintrust.
What is the cheapest LLM observability tool for a small team?
On published prices, at 300,000 traces, five engineers and about 10 KB of data per trace, Braintrust Starter ($58) and Comet Opik Pro ($89) cost the least, with PostHog ($156) and Langfuse Core ($163.30) next. Braintrust's bill scales with data ingested, so at 50 KB per trace Starter rises to about $106. Self-hosting Langfuse or Opik removes the vendor bill if you can run the infrastructure.
Next step
If you're instrumenting an agent for the first time, start with the agent evaluation regression harness workflow. Enterprise teams choosing a full agent stack should read the Enterprise AI Agent Platform Buyer's Guide.
Related Workflows
Internal Developer Productivity Agent Platform
A governed internal platform for shipping coding and ops agents to engineering teams — with shared guardrails, evals, and observability instead of shadow tools.
View workflowBrowser-Using Agent for Vendor Research & Pricing Sweeps
Agentic browser session that visits 50 vendor sites, captures pricing, terms, and feature differences, and produces a comparison sheet — overnight.
View workflowAI Virtual Ridealongs for In-Person Sales Coaching
Record in-home and in-office sales conversations so one manager can coach 25 appointments a week instead of six.
View workflowKeep Reading
Pinecone vs Weaviate vs Qdrant vs Milvus (and pgvector): Vector Database Pricing 2026
Seven vector database options priced from each vendor's own pricing page and docs, with a worked monthly bill for one million 1,536-dimension vectors and when plain pgvector is the right answer.
PostHog vs Microsoft Clarity vs Hotjar (2026): Free Session Replay, Pricing, and Consent
Microsoft Clarity, PostHog, and Hotjar (now Contentsquare) all record sessions, but they bill, retain data, and handle consent very differently. Here is what each costs for a 50,000-session small-business site, from each vendor's own pricing pages.
Vapi vs Bland vs Retell (2026): AI Voice Agent Pricing, Limits and Real Monthly Cost
Vapi, Retell and Bland price AI phone agents in three different ways. We read each vendor's own pricing page and worked out what 2,000 inbound minutes a month actually costs on each.
Found this helpful?
Get weekly AI workflow ideas in your inbox.