Langfuse vs LangSmith in 2026 - The Honest Head-to-Head

LangChain cut its old free tier, Reddit lit up, and a visible wave of users moved to self-hosted Langfuse. That migration is the whole story of this matchup in one sentence. LangSmith is the most turnkey observability you can point at a LangChain app - and it is closed-source, coupled to one framework, and roughly 25x more expensive than Langfuse at production scale.

So these two are not really competing on the same axis. LangSmith wins integration depth for LangChain code. Langfuse wins on openness, price and portability. The right pick is entirely about which of those you are optimizing for. I have added Braintrust as the third option, because a lot of people on LangSmith are really there for the evals, and if that is you, neither of the top two is the most turnkey answer.

The short version

Tool	Best for	Starting price	Self-host	License
Langfuse	Cheap self-host, framework-agnostic	Free / $29/mo	Free, near-complete	MIT
LangSmith	LangChain and LangGraph apps	$39/seat/mo	Enterprise only	Closed
Braintrust	Turnkey evals and CI regression gates	Free / $249/mo	Hybrid, Enterprise	SDKs open, backend closed

LangSmith: the deepest LangChain tracing, at a price

If you build on LangChain, LangSmith is the observability layer that just works. Add a callback and every chain, tool call and agent step shows up traced with zero extra instrumentation. No other platform gives you that depth on LangChain code, because no other platform is built by the people who ship LangChain. The eval side is strong too, and Align Evals - a workflow for calibrating your LLM judge against human scores - is genuinely useful when your automated grader disagrees with your reviewers.

It is well-funded, so survival is not the question. LangChain, Inc. raised a $125M Series B in October 2025 at a $1.25B valuation, roughly $260M total.

The gotcha is two-headed - price and lock-in. At 1M base traces a month on one seat, LangSmith runs roughly $2,514/mo, with base traces at $2.50 per 1,000 (14-day retention) and extended at $5.00 per 1,000. And you cannot self-host your way out - LangSmith is fully closed source, and self-hosting is Enterprise-only behind a sales call. The deepest value binds you to the LangChain and LangGraph stack. It is great until you scale or want out, and you cannot self-host around either problem.

Langfuse: open source that does not cripple you

Langfuse is the default answer to “what is the self-hostable alternative to LangSmith,” and it earns it. Most “open-source” observability tools give you a hobbled core with the good features gated behind a paid cloud. Langfuse does not. In the self-hosted build, only three features are enterprise-gated - tracing, evals, prompt management, human annotation and RBAC are all free under MIT. It is framework-agnostic, it runs as an OpenTelemetry backend on an OTLP endpoint, and it works with any stack rather than one ecosystem.

The economics seal it. At 1M events a month Langfuse runs about $101/mo managed, against LangSmith’s roughly $2,514/mo for comparable volume. Self-host and the per-trace cost disappears. There are real discounts too - 50% off year one for startups, up to 100% off for students.

The catch is operational, not commercial. Langfuse v3 moved from “just Postgres” to Postgres plus ClickHouse, Redis and S3-compatible storage - four services to stand up, and the migration to that architecture is where self-hosters get stuck, with GitHub threads reporting memory errors and timeouts. gRPC OTLP is not supported yet, HTTP only. And one fact to file away for a multi-year bet - Langfuse was acquired by ClickHouse in January 2026. The MIT license and self-host survived, but you are now betting on a database company’s subsidiary. If you can run the four-service stack, nothing else gives you this much for free. If you cannot, the $29/mo Core tier sidesteps it and still beats LangSmith badly.

Braintrust: if evals are the actual point

A lot of LangSmith users are really there for Align Evals, not the tracing. If regression testing is your center of gravity, Braintrust is more turnkey out of the box. Its autoevals library ships working scorers - exact match, embedding similarity, LLM-as-judge factuality - and its CI/CD quality gates can block a merge on a statistically significant regression, not just log that quality dropped. There is no per-seat charge, a nice contrast to LangSmith’s $39 a head.

The gotcha is the billing meter. Braintrust bills “processed data” in GB, counting every byte of inputs, outputs, prompts and metadata, with no hard spending cap. The $0 Starter jumps straight to $249/mo with nothing in between. Verbose agents and big RAG contexts burn the allowance fastest. Set billing alerts on day one.

Langfuse vs LangSmith: which should you pick?

• You live in LangChain and LangGraph and the bill does not scare you - stay on LangSmith. The zero-config tracing really is the deepest available for that stack, and Align Evals is a genuine reason to be here.
• You are cost-sensitive at production volume - Langfuse, no contest. The roughly 25x gap at 1M traces is the whole argument, and it has held for a while.
• You are not on LangChain at all - Langfuse. LangSmith’s premium buys you LangChain integration you are not using.
• You need free, self-hosted observability with your data on your own cloud - Langfuse, if you can run the stack. LangSmith cannot self-host below Enterprise.
• Evals and CI regression gates are the priority - Braintrust, with billing alerts on from day one.

The honest case for LangSmith: if you are deep in LangChain and your trace volume is modest, the integration is tighter than anything Langfuse gives you, and the eval tooling is first-class. The lock-in and the bill only bite when you scale or try to leave.

Want the wider field? See our LangSmith alternatives and Langfuse alternatives breakdowns, or the three-way Langfuse vs LangSmith vs Braintrust. Every price here was read from each vendor’s own materials on 26 July 2026. This category ships breaking changes monthly, so we re-verify every 30 days. The one number that decides most migrations - the roughly 25x gap at 1M traces - is not close.