LangSmith vs Helicone vs Langfuse vs PromptLayer (2026)

LangSmith, Helicone, Langfuse, and PromptLayer are production LLM observability platforms — they log API calls after your application makes them, trace multi-step chains, and track cost and latency over time. They answer "what happened when this ran in production?"

That is a different question from "is this prompt good before I ship it?" — pre-deployment testing and evaluation is the job of tools like Promptfoo, Braintrust, and Vellum, covered in Prompt Testing & Evaluation Tools 2026. Prompt version control and team sharing before deployment is covered in Best Prompt Management Platforms. The tools on this page start earning their keep the moment a prompt goes live and real users start calling it.

The overlap: all four also offer some evaluation and dataset features, because knowing a prompt broke in production is more useful if you can also score how badly. But scoring is a secondary feature here, not the primary one — the primary job is tracing, logging, and cost visibility for live traffic.

Disclosure: every vendor link on this page (LangSmith, Langfuse, Helicone, PromptLayer) points directly to that vendor’s own site. These are plain product links — none of the four currently runs a public affiliate or commission program, so there is no current affiliate relationship to disclose beyond that.

We evaluated the four tools against five criteria that matter for production tracing specifically: integration effort, self-hosting availability, framework breadth, cost transparency at scale, and eval features layered on top of logging.

LangSmith is LangChain’s own observability platform — tracing is automatic for LangChain and LangGraph applications and requires the least setup of any tool here if you already build on those frameworks. Outside the LangChain ecosystem, LangSmith still works via its SDK, but the automatic-tracing advantage disappears. Start free with the Developer tier — no card required for the first 5k traces/month.

Free Developer tier: 1 seat, up to 5,000 base traces/month, then pay-as-you-go. Plus tier: $39/seat/month, unlimited seats, up to 10,000 base traces/month included, then usage-based billing on top — $1.50 per LangChain Compute Unit (LCU) and $1.00 per LangChain Storage Unit (LSU), which meter Engine, Fleet, Deployment, and Sandbox usage. Enterprise is custom-priced with self-hosted and hybrid deployment, SSO, and RBAC.

Best team features: dataset-driven evaluation with LLM-as-judge scoring, regression testing against saved traces, and prompt hub integration so evaluation and versioning live in one place. Tradeoff: the deepest value is specific to LangChain/LangGraph — teams on a different framework get a capable but generic tracer, not the integrated experience LangChain users see.

For teams evaluating whether to standardize on LangChain in the first place, see GPT, Claude, or Gemini? How to Pick the Right Model before locking into a framework-specific observability stack.

Langfuse is MIT-licensed and free to self-host with every core feature included — there is no feature gate separating the free self-hosted version from the paid cloud tiers. Langfuse was acquired by ClickHouse, Inc. in January 2026; the product stays open source and self-hostable, and Langfuse Cloud continues running unchanged. Start free with self-hosting or the Cloud Hobby tier — no card required.

Self-hosted (Open Source): free, unlimited, MIT license, all core observability/evaluation/prompt-management features. Cloud Hobby: free, 50,000 units/month, 2 users, 30-day data retention. Cloud Core: $29/month, 100,000 units/month included, unlimited users, 90-day retention, $8 per additional 100k units. Cloud Pro: $199/month, 100,000 units/month included, unlimited users, 3-year data retention, plus a $300/month Teams add-on for SSO/RBAC. Cloud Enterprise: $2,499/month, dedicated support engineer, audit logs, custom SLA. Self-Hosted Enterprise (separately priced, bundled with a ClickHouse commercial plan) adds management APIs, project-level RBAC, and server-side data masking on top of the free self-hosted core.

Broadest framework integration of the four: OpenAI SDK, LangChain, LlamaIndex, Haystack, LiteLLM, Vercel AI SDK, Ollama, and Amazon Bedrock all have native integrations. Eval depth matches LangSmith — dataset-driven evaluation with LLM-as-judge scoring and regression datasets are both mature, not bolted-on logging extras.

Tradeoff: self-hosting means you own uptime, upgrades, and infrastructure cost — "free" software still needs someone to run it. The Cloud Hobby tier’s 2-user cap is tight for anything beyond a solo project, pushing teams to the $29/month Core tier fairly quickly once more than two people need access.

Helicone is a reverse proxy — you point your API base URL at Helicone and add one header, and every call is logged automatically, with no SDK wrapper to install. This is the fastest of the four to add to an existing codebase, since most teams already have a single place where the API base URL is configured. Start free with the Hobby tier — no card required for the first 10k requests/month.

Hobby (free): 10,000 requests/month, 1 GB storage, 7-day retention, 1 seat. Pro: $79/month, unlimited seats, 1-month retention, alerts and reports, HQL query language. Team: $799/month, SOC-2 and HIPAA compliance, 3-month retention, dedicated Slack channel. Enterprise: custom pricing, on-prem deployment, SAML SSO, custom MSA.

Because it works at the proxy layer, Helicone is provider-agnostic in a different way than LangSmith or Langfuse’s SDK integrations — it works with any OpenAI-compatible endpoint without a framework-specific adapter. Eval and testing features exist but are lighter than LangSmith or Langfuse’s dataset-driven scoring — Helicone’s core strength is logging and cost visibility, not evaluation depth.

PromptLayer is the leanest of the four — request logging and prompt versioning without a full evaluation framework, aimed at small teams that don’t need dataset-driven scoring. It wraps the OpenAI and Anthropic SDKs directly rather than working as a proxy or a deep framework integration. Start free — no card required for the first 2.5k requests/month.

Free: 5 users, 2,500 requests/month, 10 prompts, 1 workspace, 250 eval-cell executions/month. Pro: $49/month, unlimited playgrounds and workspaces, $0.003 per transaction overage. Team: $500/month, 25 users, 100,000+ requests/month, $0.002 per transaction overage, webhooks included. Enterprise: custom pricing, RBAC, deployment approvals, HIPAA with BAA, and the only self-hosted deployment option PromptLayer offers — available on GCP, AWS, and Azure at the Enterprise tier only.

Tradeoff: no self-hosted or open-source tier at any price below Enterprise — the only one of the four without one — and its framework integration list is the thinnest (OpenAI/Anthropic SDK wrappers, not the broad native-integration list Langfuse offers). Best fit is a small team that wants requests logged and prompts versioned without adopting a heavier evaluation platform.

Before wiring LangSmith, Langfuse, Helicone, or PromptLayer into a specific model provider’s calls, use PromptQuorum to dispatch one prompt to 25+ models simultaneously and confirm you picked the right provider to trace in the first place. Free tier available.

Observability tools measure what you already chose to run in production — they don’t help you pick the model. PromptQuorum answers "which model handles this prompt best?" before you commit engineering time to instrumenting production tracing for it.

No single tool wins every criterion. LangSmith goes deepest inside LangChain; Langfuse offers the most flexibility (free self-host, broadest integrations); Helicone is fastest to set up; PromptLayer is the leanest and cheapest at small scale.

Match the tool to your framework, compliance needs, and team size — not to which one has the most features on paper.

Skip all four of these tools if you have no LLM feature in production yet — pre-deployment testing and evaluation (Promptfoo, Braintrust, Vellum — see Prompt Testing & Evaluation Tools 2026) solves a real problem before you ship; observability tooling solves a problem you don’t have yet. Wiring up tracing before you have production traffic to trace is wasted setup time, and the free tiers here are generous enough that adding one later, once you actually have live traffic, costs nothing but a day of integration work.