Eval and observability platforms: Langfuse vs LangSmith vs Braintrust vs Phoenix

Benchmarks tell you how a model ranks; your production agent still fails in ways no public exam covers. An eval and observability platform is where teams watch what their agent actually did (tracing: recording every step, tool call, and token of a run) and score it continuously (evals: checks that run on those traces, in CI or on live traffic). This is also where the eval budget actually gets spent, which makes it the most-compared purchase in the whole evaluation space. Our Agent evals page covers the open-source benchmarks and frameworks one layer down.

Langfuse LangSmith Braintrust Phoenix
What it is Open-source tracing, evals, and prompt management in one platform LangChain's commercial platform for the full agent lifecycle Eval-first platform: versioned datasets, scoring, CI release gates Arize's local-first tracing and eval layer, with managed Arize AX above it
Open source? Yes: MIT core (33k stars; enterprise folders separately licensed) No: closed platform, open client SDK only No: closed platform, open SDKs only Source-available: Elastic License 2.0 (11k stars), self-hostable but not OSI open source
Self-hosting Free, first-class (Docker Compose, Helm) Enterprise-only, paid Enterprise-only, paid Free, local-first by design
Center of gravity Traces first, evals on top Lifecycle: tracing, evals, deployment, tuned for LangChain/LangGraph Scores first: "what reaches production" gates Traces and experiments, on your own machine
Free tier (checked 2026-08-12) Hobby cloud: 50k units/month, plus unlimited free self-host (pricing) Developer: 5k traces/month, 1 seat (pricing) Starter: $10 credits, 14-day retention (pricing) Self-host free with no event caps; managed AX from $0 (pricing)
Steward Langfuse, acquired by ClickHouse in January 2026 LangChain Braintrust Arize AI

No star row here because two of the four have no open-source core to count. Langfuse, Phoenix, and Opik (below) carry entries in this list's observability category.

Pick by situation

The vendors' own head-to-head pages, read against each other, agree more than you'd expect:

The fifth name that keeps appearing in these comparisons: Opik (Comet, 21.3k stars, Apache-2.0), whose whole core feature set is free to self-host under a plain open-source license; the pick when Langfuse's enterprise-folder split bothers you.

How this layer relates to the benchmarks

These platforms score your traffic continuously; benchmark frameworks like SWE-bench and inspect_ai score fixed task sets once. Mature setups use both: a benchmark to choose the base model, a platform to catch the regressions your users would otherwise find. Dev-time eval libraries (promptfoo, DeepEval, Ragas) plug into either layer; worth knowing that OpenAI acquired promptfoo in March 2026, which has non-OpenAI teams re-checking their neutrality assumptions.


Part of best-of-Agent-Harnesses. New to this decision? Start with How to pick a harness. Spot an error or a stale claim? Open an issue.

Agents can query this page's data directly: claude mcp add agent-harnesses -- uvx agent-harnesses-mcp (MCP server).