Langfuse vs LangSmith: pricing, self-hosting, lock-in

Both tools trace LLM calls, store prompts, and run evals. The differences that decide the choice are on the pricing pages and in the license, so this comparison sticks to what those pages say as of September 2026, and then explains where a tool like ClawMetry fits next to either one.

Published pricing

Langfuse Cloud

• Hobby: free, 50k units a month, 2 users, 30-day retention.
• Core: $29 a month, 100k units included, unlimited users, 90-day retention, overage $8 per 100k units.
• Pro: $199 a month, 100k units included, unlimited users, 3-year retention, SOC 2 and ISO 27001 reports.
• Enterprise: $2,499 a month, same overage, enterprise security controls.

Overage is graduated: $8 per 100k below 1M units, $7 from 1M to 10M, $6.50 from 10M to 50M, $6 above that.

LangSmith

• Developer: $0 per seat, 1 seat, up to 5k base traces a month, then pay as you go.
• Plus: $39 per seat per month, unlimited seats, up to 10k base traces a month, then pay as you go.
• Enterprise: custom pricing. Self-hosted and hybrid deployment are on this plan only.

Base traces are kept 14 days; extended traces are kept 400 days for an additional fee. The pricing page does not publish the per-trace overage rate. Third-party comparisons quote $2.50 per 1,000 base traces beyond the allowance; treat that number as theirs, not LangChain's.

The math at volume

A "unit" in Langfuse and a "trace" in LangSmith are not the same thing (a Langfuse unit is a span or event, a LangSmith trace is a whole run), so any side-by-side is approximate. With that caveat: 1M Langfuse units a month on Core is $29 plus 900k units of overage at $8 per 100k, about $101. On LangSmith Plus the per-trace charge dominates the bill at that volume, and at the third-party quoted rate 1M traces is roughly $2,500 before seats. The gap is the reason "langsmith alternative" is a search term.

Self-hosting and license

• Langfuse: MIT licensed except the ee folders, self-hostable for free on every tier, runs on Postgres, ClickHouse, Redis and S3-compatible storage. About 34k GitHub stars. Since January 2026 the team is part of ClickHouse; the repository and license are unchanged.
• LangSmith: the platform is proprietary (client SDKs are open source). Self-hosting requires the Enterprise plan.

Framework lock-in

LangSmith is LangChain's own product. If your agents are LangChain or LangGraph, tracing, Prompt Hub and annotation queues work with no assembly. Langfuse is framework-neutral by design: its SDKs are built on OpenTelemetry, spans can be exported to any OTel-compatible backend, and it lists integrations with over a hundred frameworks and providers. If you are not on LangChain, Langfuse is the shorter path.

Which one

• Pick Langfuse for open source, free self-hosting, unlimited seats on the $29 plan, or any volume past about 100k events a month.
• Pick LangSmith if you are committed to LangChain or LangGraph, want first-party tooling, and your volume stays modest.

Where ClawMetry sits

Both tools trace code you instrument. Most of the agent activity on a developer's machine today is not code you wrote: Claude Code, Codex, Cursor, OpenClaw, Gemini CLI and the rest run their own loops and write their own session files. Neither Langfuse nor LangSmith sees those unless you build a bridge.

ClawMetry works from the other side. It reads the session files those runtimes already write, needs no SDK and no code change, and shows every session, tool call, sub-agent and dollar as it happens, with detectors for runs that stall and a kill switch for the ones that should not continue. It is not a replacement for tracing inside an application you build; it is the view of everything else. Teams run it next to either of the tools above. For head-to-head detail see LangSmith vs ClawMetry and Langfuse vs ClawMetry.

Sources

• Langfuse pricing
• LangSmith pricing
• langfuse/langfuse on GitHub
• Langfuse: LangSmith alternative (Langfuse's own comparison)
• Morph: Langfuse vs LangSmith (source of the per-trace estimate)