LLM Observability Tools: The Short Answer

The best LLM observability tools in 2026 are Langfuse for framework-agnostic open-source tracing, LangSmith for teams standardised on LangChain and LangGraph, Arize Phoenix for OpenTelemetry-native instrumentation with a deep evaluation library, Braintrust for evaluation-driven development with CI gates, Helicone for proxy-based monitoring that installs in one line, and Datadog LLM Observability for organisations already running Datadog. The choice is decided by two questions, not by feature lists: which framework your application is built on, and whether your traces contain data that can leave your infrastructure.

Most teams pick an observability platform twice. The first choice is made during prototyping, when the only question is whether anyone can see what the model returned. The second is made three months into production, after a prompt change silently degraded output quality for a fortnight and nobody noticed because latency and error rates never moved. That second choice is the expensive one, and it is the one this comparison is written for.

The market has also moved underneath buyers. Langfuse was acquired by ClickHouse. In August 2026, Dynatrace announced a definitive agreement to acquire Arize, with Arize AX and Phoenix remaining available as of that announcement. Three of the platforms most teams shortlist changed ownership inside roughly twelve months, which is a reason to weight portability more heavily than any individual feature.

LLM observability tools compared

The table below compares the six platforms production teams most often shortlist. Deployment model and framework fit drive the decision far more than dashboard design.

PlatformDeploymentStrongest atBest fit
LangfuseCloud or self-hosted, MIT-licensed coreFramework-agnostic tracing, prompt versioning, production A/B testingTeams wanting open-source tracing without provider lock-in
LangSmithCloud, enterprise self-host availableDeepest LangChain and LangGraph integration, annotation queuesApplications already built on the LangChain stack
Arize PhoenixLocal-first or self-hosted, Elastic Licence 2.0OTel-native tracing, evaluation metrics library, embedding and retrieval visualisationRAG systems where retrieval quality is the failure mode
BraintrustCloudEvaluation-driven development, CI/CD quality gatesTeams that want releases blocked on eval regressions
HeliconeProxy, cloud or self-hostedOne-line install, multi-provider cost visibility, cachingVanilla API calls where install effort is the constraint
Datadog LLM ObservabilityCloudCorrelating model behaviour with infrastructure telemetryOrganisations where Datadog is already the standard

Vendors: this comparison is reviewed and republished each quarter. If a detail about your product is wrong or out of date, send us the correction and we will fix it. If you build in this category and think your product belongs on the list, tell us about it here. We review submissions on merit and disclose any paid placement on the page.

Platform-by-platform breakdown

The table sets the shape. What follows is what each platform is actually like to run, including the trade-off each one asks you to accept.

Langfuse

Langfuse became the most widely adopted open-source LLM observability platform by being framework-agnostic and genuinely self-hostable, with an MIT-licensed core. Prompt management is its strongest surface: versioning, deployment labels and production A/B testing sit alongside tracing rather than in a separate tool. Its acquisition by ClickHouse signals long-term investment in the underlying data infrastructure, which matters when trace volume grows into the hundreds of millions of spans. Pricing on Cloud is usage-based; self-hosting shifts the cost to your own infrastructure and operations time. The trade-off is that per-framework depth is broader than it is deep, so a LangChain team gets less framework-specific insight than LangSmith would give them.

LangSmith

LangSmith is LangChain's commercial platform, and the integration with LangChain and LangGraph is the deepest available because the same team builds both. Chain and graph structure appears natively in traces rather than being reconstructed from generic spans, and annotation queues make human review of production outputs a first-class workflow rather than a spreadsheet export. It is framework-agnostic in principle, but the reason to choose it is the framework coupling. The trade-off is exactly that coupling: the advantage evaporates if you migrate off the stack, which is a real risk in a category moving this fast.

Arize Phoenix

Phoenix is the OpenTelemetry-native option, released under Elastic Licence 2.0 and designed to run local-first or self-hosted. Four capabilities form a tight loop: tracing, evaluation, dataset management and a prompt playground. You trace what the application did, evaluate whether outputs met a threshold, curate the failures into a dataset, then iterate before redeploying. Its embedding and retrieval visualisations are the strongest in the field, which makes it the natural pick when the failure mode is retrieval rather than generation. In August 2026 Dynatrace announced a definitive agreement to acquire Arize, with Arize AX and Phoenix both remaining available as of that announcement. Worth tracking, though the OTel foundation means instrumentation stays portable regardless of outcome.

Braintrust

Braintrust treats evaluation as the primary object and tracing as support, which is the inverse of most platforms here. That framing suits teams practising evaluation-driven development, where a prompt change ships only after passing a scored suite in CI. If you want releases gated on quality regressions rather than reviewed after the fact, this is the platform built for it. The trade-off is that teams without an existing evaluation discipline find the model demanding, because the tool assumes you know what good looks like and can express it as a scorer.

Helicone

Helicone routes calls through a proxy, so instrumentation is a base URL change and traces start appearing in minutes with no SDK work. Multi-provider cost visibility and caching come with it, and it can be self-hosted. For a single-prompt feature or an early-stage product, it is the fastest path from nothing to visibility. The trade-off is resolution: proxy traces sit at the API-call level, not the agent-execution level, so a retry loop inside one agent step looks like several unrelated calls. Many teams run Helicone as a gateway layer and a deeper platform as the source of truth for quality.

Datadog LLM Observability

Datadog extends existing APM into LLM traffic, and its advantage is correlation: a latency spike in a generation sits next to the CPU, memory and network metrics from the same window, in a tool your on-call engineer already has open. For organisations where Datadog is the observability standard, that operational continuity often outweighs deeper AI-native features. The trade-off is depth of evaluation workflow, which is thinner than the purpose-built platforms, and cost at high trace volume.

What separates these platforms in production

The real differentiator is where each tool sits in the request path, because that determines the granularity of what it can see. A proxy captures the API call. An SDK captures the application's own execution structure.

Proxy-based versus SDK-based tracing

Helicone routes calls through its own endpoint, so instrumentation is a base URL change and traces appear immediately. The trade-off is resolution. A proxy sees a sequence of API calls; it does not see that calls three through seven were a retry loop inside a single agent step. For a single-prompt feature that distinction rarely matters. For an agent that plans, calls tools and revises, it is the whole problem.

Framework coupling

LangSmith understands LangChain internals natively, which makes debugging chains and graphs materially faster than a generic tracer. That advantage inverts if you later move off the framework. Platforms built on OpenTelemetry, particularly Arize Phoenix, keep instrumentation portable, so swapping the backend does not mean re-instrumenting the application. Given how much ownership churn the category has seen, portability is worth paying for.

A latency dashboard will tell you the model answered in 900 milliseconds. It will not tell you the answer was wrong.

Evaluation is where most implementations stall

Tracing is the easy half. The half teams underestimate is the loop back from production failures into a regression suite: capturing a bad output, turning it into a labelled test case, and blocking the next release if it regresses. Braintrust is built around that loop. Phoenix ships evaluation primitives alongside tracing. Langfuse covers scoring with strong prompt versioning attached. Whichever you choose, budget engineering time for evaluator design, because an evaluation suite that nobody curates decays into a dashboard nobody reads.

How to choose an LLM observability platform

Work through four questions in order. Most shortlists collapse to one or two candidates by the third.

1

Can traces leave your infrastructure? LLM traces contain raw prompts and completions, which in regulated environments means customer records and proprietary logic. If the answer is no, the field narrows to Langfuse, Phoenix or self-hosted Helicone immediately. Teams working under sector-specific rules should settle this alongside their wider AI strategy and governance approach rather than treating it as a tooling detail.

2

What framework is the application built on? A LangChain application gets more from LangSmith than from anything else. A LangChain-based stack and a hand-rolled one have genuinely different answers here.

3

Is the failure mode retrieval or generation? If users complain that answers are missing information rather than that answers are wrong, the problem is retrieval, and Phoenix's embedding and retrieval visualisations earn their place. That diagnosis usually points back at RAG architecture and knowledge system design rather than at the model.

4

Who owns quality after launch? If nobody owns it, no platform helps. This is an operating model question that sits closer to MLOps and ML engineering practice than to procurement.

Instrument before you need it

Retrofitting tracing onto a live agent is significantly harder than building it in, because span boundaries have to be reverse-engineered from code that was never structured to expose them. The OWASP Top 10 for LLM Applications treats monitoring gaps as a contributing factor across several of its listed risks, not as a separate concern.

Where the gap usually is

In most engagements the tool is not the bottleneck. The bottleneck is that nobody on the team has run an evaluation suite in production before, so the platform gets installed, produces traces, and then sits there. Instrumentation standards, evaluator design and the release gate are the parts that need someone who has done it.

That is the shape of work Xylity handles as a consulting-led contingent talent partner. When a team needs an engineer who has already built this loop, specialists are matched through a 4-stage consulting-led process with a 92% first-match acceptance rate across 20+ technology domains. Teams building the capability in-house usually start by adding an LLM engineer or an AI architect for the instrumentation phase, then keep ownership internally. Where the wider platform needs designing rather than staffing, that runs through our LLM application development practice, part of Xylity's broader AI consulting services.

Sector context changes the answer more than most buyers expect. A BFSI deployment carries audit obligations that make self-hosting close to mandatory, while a healthcare deployment has to treat every trace as potential patient data. Our clinical guidelines retrieval case study shows how that constraint shaped the architecture from day one.

Two related pieces cover the layers either side of this one: LLM application architecture and prompt chains for what you are tracing, and LLM cost optimisation through caching and routing for what to do once traces show where the spend is going.

Frequently Asked Questions

What is the difference between LLM observability and traditional APM?
Traditional APM measures whether a system responded and how fast. LLM observability measures whether the response was correct. APM surfaces latency, error rates and throughput. LLM observability captures the full trace of a generation, including the prompt, retrieved context, tool calls and final output, then attaches a quality score. Most teams run both, because a latency spike and a quality regression are different failures needing different signals.
Datadog LLM Observability is usually enough if the LLM feature is small and your team already lives in Datadog, because correlating model behaviour with infrastructure metrics in one place has real operational value. Teams shipping agents with multi-step reasoning generally add an AI-native platform alongside it, since agent-level trace structure and evaluation workflows go deeper in the purpose-built tools.
Self-host when traces contain regulated or confidential data. LLM traces capture raw prompts and completions, which frequently include customer records, internal documents and proprietary logic, so sending them to a third-party SaaS is a data residency decision rather than a tooling preference. Langfuse, Arize Phoenix and Helicone all support self-hosted deployment.
A proxy-based tool can be live in under an hour because it only needs a base URL change. SDK-based tracing across a multi-step agent typically takes one to two engineering weeks to instrument properly, including span naming conventions, metadata standards and a first set of evaluators. Teams building on OpenTelemetry usually recover that time later, because the instrumentation stays portable if the backend changes.

Key takeaway

Choose on deployment constraints and framework fit first, feature depth second. Given how much the category has changed hands in the past year, prefer OpenTelemetry-based instrumentation so the backend stays swappable. The best LLM observability tools in 2026 are the ones your team will still be able to move away from.

Continue building your understanding with these related resources.

5,000+specialists

Xylity matches specialists who have already run evaluation and tracing in production, not engineers learning the category on your project. The network spans 20+ technology domains, so an observability engagement can pull in data engineering or platform skills without a second procurement cycle.

See How We Work →
Best Vector Databases for Enterprise RAG in 2026

Best Vector Databases for Enterprise RAG in 2026

Best Vector Databases for Enterprise RAG in 2026 Best Vector Databases for Enterprise RAG in 2026 Best Vector Databases for ...
Best AI Governance Platforms in 2026

Best AI Governance Platforms in 2026

Skip to content Home›AI & Automation›Best AI Governance Platforms in 2026 AI & Automation10 min readAugust 2026Best AI Governance Platforms ...
Best LLM Gateway Software in 2026

Best LLM Gateway Software in 2026

Skip to main content Home › AI & Automation › Best LLM Gateway Software in 2026 AI & Automation8 min ...
Best AI Red Teaming Tools in 2026

Best AI Red Teaming Tools in 2026

Skip to main content Home › AI & Automation › Best AI Red Teaming Tools AI & Automation Best AI ...
How to Build an AI Center of Excellence in Your Organization

How to Build an AI Center of Excellence in Your Organization

Skip to content Home›AI & Automation›How to Build an AI Center of Excellence in Your Or AI & Automation12 min ...
Fine-Tuning vs RAG: Which Approach for Your LLM Application?

Fine-Tuning vs RAG: Which Approach for Your LLM Application?

Fine-Tuning vs RAG for LLM Apps: Comparison 2026 Fine-Tuning vs RAG: Which Approach for Your LLM Application? Fine-Tuning vs RAG: ...