LangSmith vs Langfuse vs Helicone: LLM observability for Next.js (2026)

Every LLM app eventually hits the same wall. It works in the demo, then something regresses in production and you cannot tell which prompt, model, or retrieval step caused it. LLM observability tools exist to give you that trace. Three names dominate the 2026 shortlist: LangSmith, Langfuse, and Helicone, and they are not interchangeable. We wired all three into a Next.js SaaS, and the honest read is that they differ most on the two things the marketing pages bury: how you actually instrument a Next.js App Router app, and who owns the trace data afterward. Here is how they compare in 2026.

Quick answer (August 2026)

Pick Helicone if you want the lowest-friction setup. It is an AI gateway, so you change one baseURL in your provider client and traces start flowing, with nothing to flush in serverless or Edge functions. Pick Langfuse if you want owned, self-hostable, OpenTelemetry-native tracing plus evaluations, with the most generous free tier (50k events per month on cloud, or unlimited self-hosted). Pick LangSmith if your Next.js app is already built on LangChain.js or LangGraph and you want the deepest evaluation, prompt, and dataset tooling in one place.

Langfuse (MIT) and Helicone (Apache-2.0) are both genuinely open source and free to self-host; LangSmith is proprietary and only self-hostable on its Enterprise plan. The real decision is your stack and your ownership stance, not raw feature count.

The 30-second version

What LangSmith is

LangSmith is the proprietary observability and evaluation platform from LangChain, the company behind the LangChain and LangGraph frameworks. You instrument a JavaScript or TypeScript app with the langsmith SDK and its traceable() wrapper, and if your app already runs on LangChain.js or LangGraph, traces appear automatically with no extra wiring. Beyond tracing, LangSmith bundles the fullest lifecycle of the three: evaluations, a prompt hub, datasets, and human annotation queues.

Pricing in 2026 (from LangChain's pricing page) is seat-plus-usage: Developer at $0 per seat (1 seat, up to 5k base traces per month then pay-as-you-go), Plus at $39 per seat (up to 10k base traces per month, unlimited seats), and a custom Enterprise plan. Overage is billed in compute and storage units (LCU at $1.50, LSU at $1.00), which is powerful but harder to forecast than a flat request count. Self-hosting is available only on Enterprise.

What Langfuse is

Langfuse is an open-source (MIT) LLM engineering platform covering tracing, evaluations, metrics, prompt management, and datasets. The repository sits at about 33.1k stars in August 2026, the core is TypeScript, and it ships both Python and JavaScript/TypeScript SDKs. It is OpenTelemetry-native: the current SDK is built on OTel, so it slots into standard tracing pipelines rather than a proprietary agent format. Langfuse is a Y Combinator W23 company, and in January 2026 it became part of ClickHouse, the analytics database Langfuse already used under the hood.

You can self-host it for free via Docker Compose, Kubernetes/Helm, or Terraform templates for AWS, Azure, and GCP. Langfuse Cloud pricing in 2026: Hobby is free (50k units per month), Core is $29 (100k units, then $8 per 100k), Pro is $199 (three-year data retention, with an optional Teams add-on at $300), and Enterprise is $2,499. It auto-instruments LangChain, LlamaIndex, and the OpenAI SDK, with manual tracing available through the SDK and decorators.

What Helicone is

Helicone is an open-source (Apache-2.0) LLM observability platform and AI gateway, sitting at about 6.1k stars in August 2026. It gives you two integration modes: the gateway, where you point your OpenAI or Anthropic client at ai-gateway.helicone.ai and add your key (one line of change), or asynchronous logging through OpenLLMetry, its OpenTelemetry integration. The platform itself is built on Next.js, Cloudflare Workers, Supabase, and ClickHouse. On top of tracing it adds response caching and routing across 100+ models with automatic fallbacks.

Self-hosting is supported via Docker Compose or Helm. Pricing in 2026: Hobby is free (10k requests per month, 1 GB, 1 seat), Pro is $79 per month (10k free requests then usage-based, unlimited seats), Team is $799 per month (SOC-2 and HIPAA, up to 5 organizations), and Enterprise is on request (on-prem deployment, SAML SSO).

The part most comparisons skip: instrumenting a Next.js app

Most head-to-heads stop at feature checklists. For a Next.js App Router app, the integration model matters more, because two of these tools are OpenTelemetry-based and one is a proxy, and that changes what you have to do inside serverless and Edge functions.

The practical takeaway: on the raw Vercel AI SDK or bare provider SDKs, Helicone is the least code to add and the safest bet inside Edge and short-lived serverless functions, because you are not responsible for flushing telemetry. Langfuse and LangSmith give you richer, structured, multi-step traces (the thing you want for TypeScript agent frameworks and RAG pipelines over a vector database), but you own the flush lifecycle in serverless. If your app already runs on LangChain.js, LangSmith's automatic tracing erases most of that friction.

Open source and ownership

Your observability data is some of the most sensitive you hold: raw prompts, user inputs, and model outputs all pass through it. That makes the license split more than a checkbox. Langfuse (MIT) and Helicone (Apache-2.0) are both genuinely open source and free to self-host, so you can keep that data inside your own infrastructure at no license cost. LangSmith is proprietary, and self-hosting is gated to its Enterprise plan. For a boilerplate you intend to ship and own, the two OSI-licensed tools keep you in control of the trace store; LangSmith trades that for the most complete evaluation suite.

Pricing for a small Next.js SaaS

The enterprise pricing tables everyone quotes (millions of traces a month) do not describe a small SaaS. Here is what the free tiers actually give a solo or small team, using each vendor's 2026 pricing.

For a small Next.js SaaS doing roughly 30,000 LLM calls a month, only Langfuse's cloud free tier (50k events) covers you at zero cost, and it is the only one of the three you can self-host free with no unit ceiling. Helicone's 10k free requests and LangSmith's 5k free traces are both fine for early testing, but you will cross them quickly in production and move to Pro ($79) or Plus ($39 per seat).

Where each one loses

No tool wins on every axis. The honest losing cases:

• Langfuse loses if you want the single deepest evaluation and prompt-lifecycle suite (LangSmith is more complete there) or a zero-config proxy (Helicone is simpler to bolt on). The ClickHouse acquisition strengthens its data backend, but adding an acquirer is a longevity question some teams will weigh.
• Helicone loses if you need rich, span-level traces of complex multi-step agents; its request-centric gateway view is thinner there than Langfuse or LangSmith, and its open-source community (6.1k stars) is smaller. A proxy in front of your LLM calls is also one more dependency in the hot path unless you self-host or lean on fallbacks.
• LangSmith loses on openness (proprietary, self-host on Enterprise only) and on cost predictability (compute and storage units are harder to forecast than a flat request count), and most of its advantage assumes you are on LangChain's JS stack. On the bare Vercel AI SDK it is more setup for less payoff than Helicone.

Bottom line

• Lowest friction on the Vercel AI SDK or bare provider SDKs, and you want a gateway with caching and routing: Helicone.
• Owned, self-hostable, OpenTelemetry-native tracing and evaluations with the most generous free tier: Langfuse.
• Already all-in on LangChain.js or LangGraph and you want the deepest eval, prompt, and dataset tooling: LangSmith.

All three are worth running. The mistake is picking on feature-list length instead of the two things that actually bite in production: how the tool instruments your Next.js app, and whether you can own the data when it matters.