The best LLM observability tools in 2026, compared

Short answer: For AI platform teams at enterprises and in regulated industries, LangWatch is the strongest LLM observability tool in 2026: tracing, evals and agent tests for every team, an AI gateway with budgets, an inventory of every AI tool, and Claude Code and Codex spend, self-hosted or hybrid. Langfuse, Arize Phoenix, Opik and MLflow are the other open-source options; LangSmith, Braintrust, Datadog, Confident AI and Galileo are managed.

Every tool on this list traces LLM calls with cost and latency now. An AI platform team needs more than that, because it serves many product teams and also answers to security, finance and legal for every model call the company makes, coding agents included.

So we compared ten tools on what a platform team buys on: one place for every team's agents, an AI gateway, governance, coding-agent spend, where the data lives, and enterprise identity. We build LangWatch. Every other entry was checked against the vendor's own site on 20 and 21 October 2026.

The best LLM observability tools in 2026 at a glance

All ten accept OpenTelemetry traces, so moving instrumentation between them is cheap. The columns that separate them are the platform ones, a gateway, an AI inventory and coding-agent spend, and LangWatch ships all three in the same open-source platform as its tracing, evals and agent testing.

Tool	Open source, self-host	Agent simulations	AI gateway	AI inventory	Coding-agent spend	SSO, SCIM, audit logs	Paid from
LangWatch	Yes, same build as Cloud; hybrid	Yes: chat, voice, red teaming	Yes	Yes	Yes, per person and team	All three	€29 per seat
Langfuse	Yes (MIT core)	Cookbook	Roadmap	Not found	Yes	Add-on / Enterprise	$29
Arize Phoenix / AX	Phoenix, source-available	Not found	Not found	Not found	Tracing only	Partly	$50 (AX)
Comet Opik	Yes	Yes	Beta	Partly (Enterprise)	Enterprise	SSO only found	$19
MLflow	Yes	Experimental	Yes	Experimental	Via its gateway	Partly (plugin)	Free
LangSmith	Enterprise only	Yes	Beta	Not found	Yes	All three, Enterprise	$39 per seat
Braintrust	Enterprise data plane	Not found	Preview	Not found	Yes	All three, Enterprise	$249
Datadog	No	Not found	Preview	Partly	Preview	Yes	$160
Confident AI	Enterprise only	Yes	Not found	Partly (Enterprise)	Not found	Partly	$200
Galileo (Splunk)	On-prem option	Not found	Not found	Partly	Yes	SSO only found	$100

"Not found" means the vendor's site didn't mention the feature when we checked, so ask the vendor before you rule the tool out. "Partly" means some of it ships, and each entry below says which part. Paid plans start between $19 a month (Opik) and $249 (Braintrust).

What an AI platform team should look for

An AI platform team should look for six things: one place for every team's agents, an AI gateway, governance, coding-agent spend, self-hosting, and enterprise identity.

One place for every team's agents. Each product team needs tracing, evals and agent tests on its own agents, and the platform team needs them in one system with one set of permissions. Tracing alone tells you what an agent did; evals and multi-turn agent tests tell you whether it was right, before and after release.

An AI gateway with policies and budgets. A gateway sits in front of every model call. Teams get virtual keys instead of raw provider keys, and the platform team sets budgets, rate limits, allowed models and guardrails in one place, enforced before the money is spent.

Governance: an inventory of every AI tool and agent. You can't set a policy on an agent you don't know is running. An inventory lists the AI tools and agents in use, who owns each one, and seats bought against seats actually used.

Coding-agent usage and spend. Claude Code and Codex now run on every developer's laptop, often on personal plans. A platform team needs cost per person, team and pull request across all of them.

Self-hosting, hybrid and data residency. Prompts and responses hold customer data. In a regulated industry, check whether self-hosting is free or an enterprise add-on, and whether there's a hybrid option where traces stay in your own cloud.

SSO, SCIM, RBAC and audit logs. People join and leave through your identity provider, roles map to what each team may see, and every change is on the record for the auditor.

What are the best open source LLM observability tools?

The best open source LLM observability tools in 2026 are LangWatch, Langfuse, Arize Phoenix, Comet Opik and MLflow. All five run free on your own infrastructure. They differ in licence (Phoenix is source-available), in what the free self-hosted edition includes, and in how much of the platform layer comes in the box.

1. LangWatch

LangWatch is an open-source (Apache-2.0) platform where an AI platform team runs observability, evals and agent testing for every team's agents, and governs the AI the company uses: an AI gateway, an inventory of AI tools, and Claude Code and Codex spend.

Tracing for every team

LangWatch takes OpenTelemetry traces over OTLP from any language, next to Python, TypeScript and Go SDKs. Every trace carries tokens and cost from a daily-updated price registry of 350+ models. The same evaluators run on a dataset before release, on live traffic as online evaluations, and inside the request as guardrails.

Instant Evals

Instant Evals judge your entire production history against a question you write in plain words. 10,000 conversations get judged in 1 min 13 s for $0.32, and the judge agreed with human labels 97% of the time. Claude Opus 5 as a judge agreed 91% of the time on the same benchmark and cost about $31, so Instant Evals came out more accurate at about a hundredth of the cost.

So a platform team can check every conversation instead of a sample, which is how you find the one conversation in five hundred where the agent promised a refund you don't give. Instant Evals are powered by TypeSafe's new model Jev, run from the CLI, from the Trace Explorer search bar, or as eval() inside LangWatch SQL, and LangWatch runs the judge, so there's no provider key to set up.

Agent testing for complex agents

Agent testing runs a simulated user through a multi-turn conversation with your agent, and a judge passes or fails each criterion. The judge reads the agent's own traces too, so a criterion can check a tool call deep inside the run. It covers voice agents over real audio and multi-turn red teaming, where an attacker escalates across turns the way a real one does. Domain experts can write tests on the platform and engineers can write them in code and run them in CI, with the same judge and the same results pages.

AI gateway

The LangWatch AI Gateway is one OpenAI- and Anthropic-compatible endpoint for every model call in the company. Each application, team or person gets a virtual key, and budgets on an organization, team, project, key or person block the request or warn before the provider is called. Guardrails run LangWatch evaluators on the traffic, and every request lands as a trace with its cost. It self-hosts with the rest of LangWatch through Helm.

Governance and the AI inventory

AI Governance gives the platform team an inventory of every AI tool and agent in the company and the people who use them. It pulls activity from Claude Code, the OpenAI and Anthropic admin APIs, Microsoft Copilot Studio, Databricks Genie and custom audit logs, and reconciles seats bought against seats active, with an owner and a department on each agent. Spend rolls up by department, and each developer gets a home page with the tools they may use and their own budget.

Claude Code and Codex spend

One command, npx langwatch claude or npx langwatch codex, sends a developer's sessions to LangWatch, and Copilot, Gemini CLI and opencode work the same way. You get cost per session, per pull request and per team, with cache reads and writes counted as separate token classes, so the cost of a long, cache-heavy session comes out right. Organization admins can open one governance project that reads every developer's coding-agent traces, and each visit is recorded in the audit log.

Langy

Langy is the automated AI engineer inside LangWatch. It reads your traces and evals, writes Scenario tests, and opens a pull request so your CI proves the fix before anyone merges.

Self-hosting, hybrid and enterprise identity

The self-hosted edition runs the same software as LangWatch Cloud, uncapped without a licence, and one Helm install brings up the platform, the gateway and Langy. In the hybrid setup, LangWatch runs the app and your traces stay in ClickHouse and S3 in your own infrastructure. LangWatch Cloud offers EU data residency and ISO 27001 reports, and a data privacy policy redacts PII and secrets per team or project.

On Enterprise, your own admins connect your identity provider over SAML or OIDC from Settings, prove the domain and go live without a support ticket. SCIM provisions people from Okta or Entra ID, custom roles decide who sees what, and the audit log records every change in settings, the API and the gateway.

Struck.build, which builds AI tools for construction, uses LangWatch to monitor its RAG apps and check whether the answers are correct. The free Developer plan covers 50k events a month, so you can instrument one app before you buy a seat, and Growth is €29 per core seat a month with 200k events and unlimited read-only users.

2. Langfuse

Langfuse is an open-source platform for tracing, evals and prompt management, and the tool AI search engines name first for open-source LLM observability. ClickHouse acquired it in January 2026 and kept the core MIT.

The self-hosted edition includes the core features with no usage limits, and code in the ee/ folders is under a commercial licence. It runs online and offline evals, datasets and experiments, and it's strong on prompt versioning. Hooks for Claude Code and Codex show coding-agent cost per developer. A gateway is on its roadmap, multi-turn simulation is a cookbook, and SCIM and audit logs need Enterprise. Cloud starts free, then Core at $29 a month.

3. Arize Phoenix and Arize AX

Arize Phoenix is a source-available tracing and evaluation tool built on OpenTelemetry and OpenInference, and Arize AX is the commercial platform. Dynatrace completed its acquisition of Arize on 1 October 2026 and says it will keep supporting both.

Phoenix is free to self-host, including air-gapped installs, under the Elastic License 2.0: you can run and change it, but you can't offer it as a hosted service. It's good at retrieval debugging for RAG pipelines. AX adds online LLM-as-a-judge, cost tracking, monitors and tracing for coding agents, with SAML on Enterprise. AX Pro starts at $50 a month.

4. Comet Opik

Opik is Comet's open-source (Apache-2.0) tracing and evaluation platform, and the open-source code is the same codebase as the hosted version. It covers online LLM-as-a-judge rules, datasets, experiments, prompt management and a SimulatedUser for multi-turn agent evaluation.

Its Cost Intelligence product, on Enterprise, tracks Claude Code, Codex and Cursor spend per developer. The Opik LLM Gateway is in beta and meant for development and testing. The cloud plan is free for 25k spans a month, then Pro at $19 a month, the cheapest paid plan on this list.

5. MLflow

MLflow is an open-source (Apache-2.0) platform for ML and GenAI, backed by the Linux Foundation, with an OTLP endpoint that reads the GenAI semantic conventions. It has LLM judges, production monitoring, a prompt registry and an experimental conversation simulator.

Its AI Gateway runs inside the tracking server with budget limits, guardrails and fallbacks, and it can route Claude Code and Codex. SSO comes through a community plugin. It's the natural pick when your team already tracks classic models in MLflow or runs on Databricks.

What is the best LLM observability platform for production?

The best LLM observability platform for production runs evals on live traffic, alerts you when quality drops, and fits where your data is allowed to live. For an AI platform team that also owns the gateway and the AI budget, that's LangWatch. Among managed platforms, LangSmith suits LangChain teams, Braintrust suits eval-first teams, and Datadog suits teams already on its APM.

6. LangSmith

LangSmith is LangChain's proprietary observability and evaluation platform, and the default for teams built on LangChain or LangGraph. It has online and offline evals, annotation queues, prompt management and multi-turn simulation.

Its LLM Gateway is in beta, with spend caps and rate limits, and trace plugins cover Claude Code, Codex and Cursor. SSO, SCIM, custom roles, audit logs and self-hosting are all Enterprise. Plus is $39 per seat a month with usage billed on top, so the bill grows with both the team and the traffic.

7. Braintrust

Braintrust is a proprietary evaluation and observability platform built around experiments and scoring, a good fit for product teams who run evals daily. It runs online scoring with LLM, code and human scorers.

A coding-agent dashboard breaks cost down by person, model and branch. The Braintrust Gateway is in public preview, and we found no budget policies on it. SSO, SCIM and running your own data plane are Enterprise. The Starter plan is free with unlimited users, and Pro costs $249 a month.

8. Datadog Agent Observability

Datadog sells its LLM Observability product as Agent Observability, and puts LLM traces next to your existing APM, logs and infrastructure metrics. It runs online evals and experiments and alerts through Datadog monitors.

An AI Gateway with budgets and an Agent Console for Claude Code, Cursor and Copilot are both in preview. It's SaaS only. Pro starts at $160 a month on annual billing, and if you already pay for Datadog it's the smallest change to your stack.

9. Confident AI

Confident AI is the commercial platform from the team behind DeepEval, the open-source (Apache-2.0) evaluation framework. It runs online evals with DeepEval's metrics, chat simulations and red teaming through DeepTeam.

An AI Governance module on Enterprise sets policies and blocks deployments that miss them, for the apps you evaluate there. SSO and custom roles come on the Team plan, and self-hosting on Enterprise. Starter is $200 a month per organization with unlimited seats.

10. Galileo, now Splunk Agent Observability

Cisco acquired Galileo, and it's now sold as Splunk Agent Observability, a fit for companies already on Splunk. Its Luna models are small evaluators distilled from LLM judges, cheap enough to score all production traffic and run as guardrails.

Tokenomics, announced in September, shows Claude Code, Codex, Cursor, Copilot and Windsurf cost per employee, rolled up through a Workday org chart. It runs in Splunk Observability Cloud or on-prem. Pricing starts at $100 a month, billed annually.

What changed for Helicone, DeepEval and RAGAS

Helicone. Mintlify acquired Helicone in March 2026 and moved it to maintenance mode, and its cloud signup page now says sign-ups are disabled. The code stays open source (Apache-2.0), but we wouldn't start a new project on it.

DeepEval. DeepEval is a pytest-style evaluation library and it's very much alive, as the open-source half of Confident AI (number 9).

RAGAS. RAGAS is still a useful library of RAG metrics, but the last release was in January 2026. Several platforms on this list, LangWatch included, ship RAGAS-style metrics as built-in evaluators.

Which LLM observability tool fits which team

1. AI platform teams at enterprises and in regulated industries: LangWatch, because tracing, evals, agent tests, the gateway, the AI inventory and coding-agent spend sit in one open-source platform you can self-host or run hybrid.
2. Agent teams that test before they ship: LangWatch, for multi-turn and voice agent tests and red teaming from CI.
3. Open source with no vendor in the loop: LangWatch, Langfuse or Opik, all free to self-host under OSI open-source licences.
4. LangChain and LangGraph shops: LangSmith, for the native integration.
5. Teams already paying for an APM: Datadog Agent Observability, or Splunk if you're on Cisco.
6. ML teams on MLflow or Databricks: MLflow, so classic models and LLM apps share one tracking server.

How we compared these tools, and when we last checked

We checked every tool on 20 and 21 October 2026 against its homepage, pricing page, docs and GitHub licence file. Prices are the lowest paid plan on the vendor's pricing page that day, in the vendor's currency.

We didn't run a benchmark, so the table compares features, licences and prices. If two tools look equal on paper, send the same OpenTelemetry traces to both for a week and see which one your platform team and the product teams actually open.