Best for: Teams with strict data residency requirements or those who want full control without vendor lock-in.
Pricing: Free self-hosted. Cloud starts at 0ドル for 50K observations/month.
2. LangSmith -- Best for LangChain Teams
LangSmith is the observability layer built by the LangChain team, and it shows. Tracing LangChain and LangGraph workflows is nearly zero-config, and the Prompt Hub plus dataset-driven evaluation workflows are mature and well-documented.
Key strength: Deepest integration with the LangChain/LangGraph ecosystem. If you're already using LCEL, tracing just works.
Key weakness: Vendor lock-in. If you ever move off LangChain, you lose most of the value. Non-LangChain tracing works but feels bolted on.
Best for: Teams already committed to the LangChain stack who want tracing and evals in one place.
Pricing: Free for 5K traces/month. Plus tier at 39ドル/seat/month.
3. Helicone -- Easiest Setup
Helicone uses a proxy-based approach: swap your OpenAI base URL and you're logging traces in under 2 minutes. No SDK, no code changes. Their cost analytics dashboard covers 100+ models and gives you instant visibility into spend by model, user, or feature.
Key strength: Fastest time-to-value. 99.99% uptime SLA and a proxy architecture that requires zero code instrumentation.
Key weakness: Request-level tracing only. You won't get the span-level granularity that SDK-based tools offer for complex chains or agent loops.
Best for: Teams that want cost visibility and basic tracing without touching their codebase.
Pricing: Free for 10K requests/month. Pro starts at 79ドル/month.
4. Braintrust -- Best for Evaluation-First Teams
Braintrust puts evaluation at the center. Their CI/CD quality gates can block deployments when quality metrics regress, and real-time dashboards flag hallucinations as they happen. If your team treats AI output quality like test coverage, Braintrust speaks your language.
Key strength: CI/CD-integrated eval gates that enforce quality thresholds before code ships.
Key weakness: Higher price point at 249ドル/month. The eval-first approach also means tracing and logging feel secondary to the scoring workflow.
Best for: Teams where AI output quality is mission-critical and regressions need to be caught before production.
Pricing: Free for 1M trace spans. Pro at 249ドル/month.
5. Arize Phoenix -- Best Free Self-Hosted
Arize Phoenix is open-source under Elastic 2.0 and comes with embedded drift detection, RAG quality metrics, and retrieval visualizations out of the box. It's particularly strong at catching silent model degradation -- the kind where outputs slowly get worse and nobody notices.
Key strength: Drift detection and RAG-specific quality plots that no other free tool matches.
Key weakness: Less polished UI than commercial options, and the Elastic 2.0 license is more restrictive than MIT for some enterprise use cases.
Best for: Teams running RAG pipelines who need quality monitoring without a SaaS bill.
Pricing: Free and unlimited when self-hosted. Cloud pricing available.
6. Datadog LLM Observability -- Best for Enterprise
Datadog LLM Observability plugs directly into the APM, logs, and metrics you already have. Built-in safety detection covers hallucinations, PII leakage, and bias. The value prop is simple: one pane of glass for your entire stack, LLMs included.
Key strength: Unified observability. Correlate LLM traces with infrastructure metrics, error rates, and deployment events in one dashboard.
Key weakness: Enterprise pricing and complexity. If you're a startup without existing Datadog, the overhead isn't worth it.
Best for: Organizations already running Datadog that want to add LLM monitoring without adopting another vendor.
Pricing: Bundled with Datadog plans. Contact sales for LLM-specific pricing.
7. Nebula -- Best for AI Agent Teams
Nebula isn't a standalone observability platform -- it's an AI agent execution platform with tracing built in. You get agent-level tracing across multi-agent workflows, three-layer safety checks on all write actions, and action labeling that distinguishes read vs write operations automatically.
Key strength: Observability is embedded in the agent runtime itself. No separate instrumentation needed for agent workflows.
Key weakness: Not a dedicated monitoring tool. If you need deep span-level tracing across arbitrary LLM calls outside of agent workflows, a purpose-built tool like Langfuse or Helicone is a better fit.
Best for: Teams already orchestrating AI agents who want built-in tracing without bolting on a separate observability stack.
Pricing: Free tier available with generous limits.
Verdict
There's no single winner here -- the right tool depends on your stack and what you care about most. Open-source loyalists should start with Langfuse. Teams that want instant setup and cost visibility should try Helicone. If you're deep in the LangChain ecosystem, LangSmith is the natural choice. Enterprise orgs on Datadog should just turn on the LLM module. Eval-obsessed teams will love Braintrust's quality gates. RAG-heavy workloads get the most from Arize Phoenix. And if you're running multi-agent workflows, Nebula's built-in tracing saves you from stitching together yet another tool.
Pick the one that fits where you are today -- you can always swap later.