Langfuse vs LangSmith vs Braintrust in 2026 - Pick by What You Actually Need

Langfuse, LangSmith and Braintrust are the three names that come up in almost every LLM observability shortlist. They overlap enough to be confusing and differ enough that picking wrong costs you real money or real lock-in.

Here is the one-line frame. Langfuse is the cheap, open-source, framework-agnostic default. LangSmith is the turnkey choice if you live in LangChain, at a bill that explodes at scale. Braintrust is the eval-first platform built to block bad merges. Match the tool to your actual center of gravity and the choice gets easy.

The short version

	Langfuse	LangSmith	Braintrust
Pricing model	Billable units, $8/100k	Base traces, $2.50/1k	Processed data by the GB
Cost at scale	~$101/mo at 1M events	~$2,514/mo at 1M traces	Depends on trace verbosity
Self-host	Free, MIT, near-complete	Enterprise only, closed	Hybrid VPC, Enterprise only
Center of gravity	Open-source observability	LangChain tracing	Evals and regression gates
Starting paid tier	$29/mo	$39/seat/mo	$249/mo

Langfuse: the cheap open-source default

Langfuse is the open-source default for LLM observability, and it earns the title on two facts.

First, self-hosting gets you the real product. Only three features are enterprise-gated in the self-host build - Organization Creators, the Instance Management API, and UI customization. Tracing, evals, prompt management, human annotation and RBAC are all free under MIT. Second, the economics - at 1M events a month it runs about $101/mo managed, and self-hosting removes the per-trace cost entirely. It is framework-agnostic, so it works with any stack, and it runs as an OpenTelemetry backend on an OTLP endpoint.

The catch is operational. Langfuse v3 needs Postgres plus ClickHouse, Redis and S3-compatible storage - four services - and the migration to that architecture is where self-hosters get stuck, with GitHub threads reporting memory errors and timeouts. OTLP is HTTP only, no gRPC yet. And it is now a ClickHouse subsidiary after the 16 January 2026 acquisition, which is worth filing away for a multi-year bet - the MIT license and self-host survived. If you cannot spare the ops capacity, the $29/mo Core cloud tier sidesteps the whole thing and still beats LangSmith badly.

LangSmith: turnkey for LangChain, brutal at scale

LangSmith is the most turnkey observability you can point at a LangChain app. Add a callback and every chain, tool call and agent step shows up traced with zero extra work. Nothing touches it for depth on LangChain and LangGraph code, because nobody else ships LangChain. The eval side is strong too, and Align Evals - calibrating an LLM judge against human scores - is genuinely useful when your automated grader disagrees with your reviewers. It is backed by a company that raised roughly $260M and is valued at $1.25B, so survival is not the question.

Two things keep it from being the default. The bill runs roughly $2,514/mo at 1M base traces on one seat - about 25x Langfuse - 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. It accepts OpenTelemetry as a receiver, so you are not forced onto the LangChain SDK, but the zero-config magic only shows up when you use it. When LangChain cut its old free tier, Reddit threads pushed a visible wave of users toward self-hosted Langfuse for exactly this reason. Live in LangChain and unbothered by the bill - stay. Otherwise, the alternatives are cheaper.

Braintrust: evals are the whole point

Braintrust is the one where evals are the center of gravity, not an add-on. Human review, automated scorers, LLM-as-judge, tracing, dataset management and CI/CD quality gates all share one system. The autoevals library ships working scorers out of the box, and the CI gates can block a merge on a statistically significant regression - actually stopping the bad code, not just logging that quality dropped. There is no per-seat charge either, so users are unlimited on every tier.

The gotcha is the billing meter. Braintrust bills “processed data” in GB, counting every byte of inputs, outputs, prompts and metadata, and there is no hard spending cap. Verbose multi-step agents and big RAG contexts - the workloads that most need observability - burn the allowance fastest. The $0 Starter jumps straight to $249/mo Pro with nothing in between. Set billing alerts on day one. And self-host is hybrid-VPC only and Enterprise-only, with the Brainstore backend closed, so you never fully own the stack.

Which should you pick?

• You want the best all-round platform for the least money - Langfuse, self-hosted if you can run the four-service stack, or the $29/mo Core cloud if you cannot. Roughly 25x cheaper than LangSmith at scale.
• You live in LangChain and LangGraph and the bill does not scare you - LangSmith. The zero-config tracing is genuinely the deepest available for that stack, and Align Evals is a real reason to be there.
• Regression testing and blocking bad merges is your priority - Braintrust, with billing alerts on from day one. Nothing else here is this turnkey for CI quality gates.
• Free open-source self-hosting is a hard requirement - Langfuse is the only one of the three that delivers it.
• You are cost-sensitive at production volume - anything but LangSmith. Langfuse for pure cost, Braintrust if evals justify the spend.

The honest close - most teams that think they need LangSmith actually need Langfuse plus a bit of setup, and they save an order of magnitude doing it. But if you genuinely live in LangChain, the integration is worth paying for, and if regression gates are your bottleneck, Braintrust earns its keep. Pick by your center of gravity, not the feature checklist. Prices here were read from each vendor’s own pages, and this category ships breaking changes monthly, so re-verify before you commit.