LangSmith Trial

Overview

LangSmith provides tracing, datasets, evaluations, and prompt hub capabilities for LangChain and compatible stacks (LangSmith).

Trial when multiple squads need shared LLMOps workflows. Avoid duplicating telemetry if OpenTelemetry GenAI conventions already feed your central observability platform unless LangSmith fills an eval gap.

Adoption Signals

• LangChain raised $125M at a $1.25B valuation in October 2025, led by IVP with CapitalG and Sapphire Ventures joining; the announcement names Replit, Clay, Harvey, Rippling, Cloudflare, Workday, and Cisco as users and reports that monthly LangSmith trace volume grew 12x year over year (LangChain Series B announcement, Fortune).
• In the same announcement, LangGraph Platform was renamed LangSmith Deployment, making LangSmith the umbrella for observability, evaluation, and agent hosting (LangChain Series B announcement).
• Klarna uses LangSmith for observability, test-driven development with LLM evaluations, and prompt optimization on the AI assistant that serves 85 million active users, alongside LangGraph for the agent itself (Klarna customer story).
• OpenTelemetry ingestion through an OTLP endpoint is documented, including a Collector fan-out pattern that routes the same spans to LangSmith and other observability backends, with attribute mapping for GenAI semantic conventions, OpenInference, and Traceloop (Trace with OpenTelemetry).
• EU data residency is offered on all plans at the same price as the US region (Regions FAQ).
• Self-hosted LangSmith has been tested on GKE, EKS, AKS, and OpenShift 4.14+ (Self-hosting on Kubernetes).

Risks

• Price and retention changes. A third-party pricing tracker recorded the July 2026 switch from per-1,000-trace dollar rates to LCU/LSU units, which it calculates as doubling base-trace cost to $5 per 1,000 (UsagePricing); the official pricing page now bills in LCU ($1.50) and LSU ($1.00) units (LangSmith pricing), and the changelog confirms that extended SaaS retention is now capped at 180 days (LangSmith changelog). Budgets and audit-retention plans built on 2025 figures need rechecking.
• Sensitive data in traces. Traces contain full prompts, retrieved documents, and tool outputs. The 2024 "AgentSmith" flaw (CVSS 8.8), found by Noma Security, let a malicious agent shared on LangChain Hub route users' prompts and OpenAI API keys through an attacker's proxy; it was patched in November 2024 (The Hacker News). Do not use shared public Hub artefacts without review, and set up PII redaction and access policies before you send production traffic.
• No region switch. An organization cannot currently move between the US and EU regions (Regions FAQ), so choosing the wrong region at sign-up means a migration later.
• Self-hosting weight. Running it yourself requires an Enterprise licence and a Kubernetes deployment with PostgreSQL, Redis, and ClickHouse (Self-hosting on Kubernetes), which is a considerable operations commitment for one observability tool.
• Platform gravity. As deployment, evaluation, and prompt management move into one commercial product, datasets, evaluators, and prompts become harder to take elsewhere; OTLP ingestion covers traces but not those assets.

Recommendation

Trial LangSmith when a team already builds on LangChain or LangGraph and needs shared tracing, evaluation datasets, and human annotation across several squads without building its own evaluation tooling. For European companies, create the organization in the EU region from the start, sign the DPA, and define which fields are redacted or excluded before production traces flow. Instrument through OpenTelemetry rather than only the LangSmith SDK, so the same spans can also reach your central observability stack and a later switch stays possible.

Do not choose it when policy requires the tool to run on your own infrastructure but there is no Enterprise budget, or when traces must be kept for longer than 180 days in the SaaS product; Langfuse or MLflow are the more suitable options there. Before rollout, model the cost from actual trace volume under the current LCU/LSU pricing and set sampling rules, and keep evaluation datasets and prompts exported to version control so they do not live only in LangSmith.

Sources

• LangSmith
• LangSmith docs
• Observability
• Evaluation
• LangSmith pricing
• LangSmith changelog
• Trace with OpenTelemetry
• Regions FAQ
• Self-hosting on Kubernetes
• LangChain Series B announcement
• Fortune: LangChain becomes a unicorn
• Klarna customer story
• UsagePricing: LangSmith price change July 2026
• The Hacker News: AgentSmith vulnerability