Best software · Developer & data
10 Best LLM Observability Tools of 2026
This list is for AI engineers and platform teams who need to trace, evaluate and debug LLM applications and agents in development and production. It covers open-source platforms you can self-host (Langfuse, Arize Phoenix, Opik, Laminar), hosted tracing and eval platforms (LangSmith, Braintrust, W&B Weave, HoneyHive), an AI gateway with built-in logging (Helicone) and the LLM module of a full observability suite (Datadog). We compared documented features and published prices for 10 tools. Scores weigh features at 40%, ease of use at 30% and value at 30%.
By the Mojolin research desk · Updated November 9, 2026 · 10 tools compared · Prices checked on vendor pages, latest November 9, 2026
Questions or corrections: editorial@mojolin.com
Mojolin is self-funded How we score
Langfuse is our top pick for most AI engineering teams: it is open source and self-hostable, covers tracing, prompt management and online and offline evals, and its cloud Core plan costs $29 per month with unlimited users. LangSmith is the runner-up and the natural choice for LangChain and LangGraph users, with SDKs for Python, TypeScript, Go and Java at $39 per seat. Braintrust is the strongest pick when evaluation is the center of your workflow, though Pro starts at $249 per month. Opik and Arize Phoenix are the best open-source alternatives, and teams already paying for Datadog can add LLM spans to the dashboards they have.
Top picks
Best overall
Langfuse
9.1 / 10
AI engineering teams that want a full open-source stack they can self-host, or a low-cost cloud plan with unlimited users.
Read the review →Runner-up
LangSmith
9.0 / 10
Teams building agents on LangChain or LangGraph, and polyglot teams that need SDKs for Python, TypeScript, Go and Java.
Read the review →Also great
Braintrust
8.8 / 10
AI product teams that treat evals as a release gate and want engineers and product managers in the same tool.
Read the review →Comparison table
| # | Tool | Pricing | Score | |
|---|---|---|---|---|
| 1 | Langfuse | AI engineering teams that want a full open-source stack they can self-host, or a low-cost cloud plan with unlimited users. | Free Hobby plan with 50k units per month; Core from $29 per month with 100k units and unlimited users | 9.1 |
| 2 | LangSmith | Teams building agents on LangChain or LangGraph, and polyglot teams that need SDKs for Python, TypeScript, Go and Java. | Free Developer plan for 1 seat with 5k base traces per month; Plus from $39 per seat per month with 10k base traces | 9.0 |
| 3 | Braintrust | AI product teams that treat evals as a release gate and want engineers and product managers in the same tool. | Free Starter plan with 1 GB of processed data and 10k scores; Pro from $249 per month with 5 GB and 50k scores | 8.8 |
| 4 |  Arize Phoenix | Engineers who want a local-first open-source tool on OpenTelemetry, with a managed upgrade path. | Phoenix is free to self-host; Arize AX Free includes 25k spans, AX Pro costs $50 per month with 50k spans | 8.7 |
| 5 |  Opik | Teams that want open-source tracing plus pass/fail test suites, with the cheapest paid cloud plan on this list. | Free open-source version and Free Cloud plan with 25k spans; Pro Cloud from $19 per month with 100k spans | 8.6 |
| 6 |  W&B Weave | ML and AI teams already using Weights & Biases for experiment tracking that want agent tracing in the same workspace. | Free plan with 1 GB of Weave ingestion per month; Pro starts at $60 per month billed monthly | 8.4 |
| 7 |  Datadog LLM Observability | Teams already running Datadog that want LLM traces next to their APM, logs and infrastructure data. | Free tier up to 40K LLM spans per month; Pro from $160 per month with 100K LLM spans, annual pricing | 8.3 |
| 8 |  Helicone | Teams that want logging, caching and provider fallbacks from a one-line gateway integration. | Free Hobby plan with 10,000 requests; Pro from $79 per month with unlimited seats plus usage | 8.2 |
| 9 |  Laminar | Teams running long agent workflows that want automatic failure detection across many runs. | Free plan with 1 GB of data; Starter from $30 per month with 3 GB and unlimited seats | 8.1 |
| 10 |  HoneyHive | Enterprises in regulated industries that need audit-ready evidence and flexible hosting for agent monitoring. | Free Developer plan with 10,000 events per month and up to 5 users; Enterprise pricing on request | 8.0 |
Scoring weights
Every tool gets three editorial scores out of 10, from vendor documentation, product pages and the published price. The overall score is the weighted average below, the same method on every list. Full method
The full ranking
Rank 1 · Best overall
Langfuse
Open-source AI engineering platform for tracing, prompt management and evaluation of LLM apps and agents, available as a cloud service or self-hosted.
langfuse.com9.1
out of 10
Best for: AI engineering teams that want a full open-source stack they can self-host, or a low-cost cloud plan with unlimited users.
Langfuse covers the full loop of LLM engineering in one open-source product. Traces capture every LLM call plus retrieval, embeddings and API calls, group multi-turn chats into sessions and can show agents as graphs, with cost and token breakdowns on each generation. Instrumentation works through Python and JavaScript SDKs, OpenTelemetry, 100+ integrations or a gateway such as LiteLLM. Prompt management handles versions, release labels and caching, and the evaluation side supports datasets, experiments, LLM-as-judge scores, user feedback and human annotation queues.
The cloud Hobby plan is free with 50k units, 30 days of data and 2 users. Core costs $29 per month for 100k units, 90 days of data and unlimited users, and Pro at $199 adds 3 years of history, higher rate limits and SOC 2 and ISO 27001 reports. Enterprise SSO sits in a $300 Teams add-on. Teams that cannot send traces out can self-host from the same codebase.
Standout feature
The same open-source codebase runs in the cloud and on your own servers, so you can start hosted and move in-house later.
Use cases
Startup AI teams
Trace a RAG chatbot and track cost per user
Outcome: Expensive or slow requests are traced to the exact retrieval step
Regulated companies
Self-host Langfuse on Kubernetes
Outcome: Prompt and response data stays inside their own cloud account
Prompt engineers
Version prompts and compare them in experiments
Outcome: A prompt change ships only after it beats the current version on a dataset
Key features
- Traces of LLM and non-LLM calls, sessions, user tracking and agent graphs
- Native Python and JavaScript SDKs, OpenTelemetry and 100+ framework integrations
- Prompt versioning, release labels, caching and a playground
- Online and offline evaluation with datasets, LLM-as-judge and human annotation queues
- Self-hosting guides for Docker Compose, Kubernetes and Terraform
Pros
- Fully open source and self-hostable
- Core plan includes unlimited users for $29 per month
- Cloud data regions in the US, EU and Japan
Cons
- Hobby plan keeps data for 30 days and allows 2 users
- Enterprise SSO and fine-grained RBAC need the $300 per month Teams add-on
- Usage beyond the included units costs $8 per 100k units
Rank 2 · Runner-up
LangSmith
LangChain's agent engineering platform for tracing, monitoring, evaluation and deployment of LLM applications built with any framework.
langchain.com9.0
out of 10
Best for: Teams building agents on LangChain or LangGraph, and polyglot teams that need SDKs for Python, TypeScript, Go and Java.
LangSmith is the observability and evaluation layer from the team behind LangChain and LangGraph, but it traces apps built with the OpenAI SDK, Anthropic SDK, Vercel AI SDK, LlamaIndex and others too. Traces show each agent step with message threading for multi-turn chats, and SmithDB, its trace store, supports full-text search and JSON key-path filtering. Monitoring dashboards cover tokens, P50 and P99 latency, errors, cost and feedback, with webhook and PagerDuty alerts. Insights cluster traces automatically to find common behaviors and failures, and evaluation covers online and offline evals, datasets, annotation queues and a prompt playground.
The free Developer plan gives one seat and 5k base traces per month. Plus costs $39 per seat per month with 10k base traces, unlimited seats and access to Deployment and Engine, then pay as you go. Enterprise adds self-hosted or hybrid deployment, custom SSO and an SLA. Seat pricing makes it costlier than Langfuse for larger teams.
Standout feature
Insights cluster production traces automatically, so teams see common agent behaviors and failure modes without writing queries.
Use cases
LangGraph developers
Debug a multi-step agent run step by step
Outcome: The tool call that caused a bad answer is found in minutes
Platform teams
Alert on cost and latency through PagerDuty
Outcome: On-call engineers hear about a runaway agent before the bill does
Product teams
Run online LLM-as-judge evals on production traffic
Outcome: Quality drops after a model change are visible within a day
Key features
- Native tracing for popular agent frameworks plus OpenTelemetry SDKs for Python, TypeScript, Go and Java
- Dashboards for token usage, P50 and P99 latency, error rates, cost and feedback scores
- Online LLM-as-judge and code evals, offline evals, datasets and annotation queues
- Insights that cluster traces to surface usage patterns and failure modes
- Webhook and PagerDuty alerts when metrics cross thresholds
Pros
- Works with any framework, not only LangChain
- Free Developer plan includes 5k base traces each month
- Deployment, sandboxes and an LLM gateway are available on the same account
Cons
- Per-seat pricing grows with team size
- The free plan is limited to one seat
- Self-hosted and hybrid deployment require Enterprise
Rank 3 · Also great
Braintrust
AI observability and evaluation platform with tracing, experiments, scorers and the Loop agent for improving prompts and datasets.
braintrust.dev8.8
out of 10
Best for: AI product teams that treat evals as a release gate and want engineers and product managers in the same tool.
Braintrust is built around evaluation. You run experiments against versioned datasets, compare prompts and models side by side and score outputs with LLM judges, code or human reviewers, then turn real production traces into regression datasets with one click. The observability side ingests agent traces, searches across long prompts and tool calls through its Brainstore database and tracks latency, cost and quality live. The Loop agent can generate better prompts, scorers and datasets from a plain description, and Patterns flags recurring issues in production traces. SDKs cover Python, TypeScript, Go, Ruby and C#, and an MCP server lets coding agents query logs and run evals.
Starter is free with $10 of model credits, 1 GB of processed data, 10k scores and 14-day retention. Pro costs $249 per month with $100 credits, 5 GB, 50k scores, 30-day retention, custom charts and RBAC. Both include unlimited users, which offsets the high Pro price for bigger teams.
Standout feature
Production traces become eval datasets with one click, so regression tests come from real failures.
Use cases
AI product teams
Gate each prompt change behind an experiment
Outcome: Regressions are blocked before release
Support automation teams
Build custom trace views for reviewing conversations
Outcome: Domain experts annotate outputs without frontend work
Engineers using coding agents
Query logs and run evals through the MCP server
Outcome: Debugging happens from the IDE
Key features
- Real-time inspection of prompts, responses and tool calls across millions of logs
- Experiments against versioned datasets with LLM, code and human scoring
- Loop agent that generates prompts, scorers and datasets
- One-click conversion of production traces into eval datasets
- SDKs for Python, TypeScript, Go, Ruby and C#, plus an MCP server
Pros
- Unlimited users, projects and experiments on every plan
- Strong side-by-side comparison of prompts and models
- 6 to 12 months of Pro free for qualifying startups
Cons
- Pro is the most expensive self-serve entry on this list at $249 per month
- Starter keeps data for only 14 days
- SAML SSO and on-prem deployment require Enterprise
Rank 4
Arize Phoenix
Open-source tracing and evaluation tool from Arize, with Arize AX as the managed platform for teams that outgrow it.
arize.com8.7
out of 10
Best for: Engineers who want a local-first open-source tool on OpenTelemetry, with a managed upgrade path.
Arize Phoenix is an open-source tool for tracing, evaluating and iterating on LLM apps and agents. It records every step an agent takes, including prompts, retrievals and tool calls, and lets you annotate outputs with human review or LLM-as-judge evals, build datasets from traces and run experiments in a prompt IDE. It is built on OpenTelemetry and the OpenInference spec, and you can start it locally with one command, run it in Docker or deploy it on Kubernetes with Helm.
Arize AX is the managed platform. AX Free includes 25k spans, 1 GB of ingestion and 15-day retention; AX Pro costs $50 per month for 50k spans, 10 GB and 30 days. Both have unlimited users, evals and experiments, plus multi-modal tracing, session evals and the Alyx agent with a monthly issue quota. Phoenix uses the ELv2 license, which allows self-hosting but restricts offering it as a managed service.
Standout feature
Phoenix starts locally with a single command, so tracing works on a laptop before any account or cloud setup.
Use cases
Solo AI engineers
Run Phoenix locally while building a RAG pipeline
Outcome: Retrieval and prompt problems are visible before deploy
Data science teams
Score outputs with LLM-as-judge evals on traces
Outcome: Hallucination and relevance issues are tracked over time
Growing teams
Move from Phoenix to Arize AX Pro
Outcome: Managed tracing with unlimited users for $50 per month
Key features
- Tracing of prompts, retrievals, tool calls and outputs on native OpenTelemetry
- Evals with LLM-as-judge and human annotations
- Datasets from traces, experiments and a prompt IDE
- Runs locally, in Docker or on Kubernetes with Helm
- Arize AX adds session evals, agent trajectory views, custom monitors and the Alyx agent
Pros
- Phoenix can run on a laptop with a single command
- Unlimited users and evals on every AX plan
- Vendor agnostic across models, frameworks and languages
Cons
- Phoenix is ELv2 licensed rather than a permissive open-source license
- AX Free and Pro keep data for only 15 and 30 days
- Self-hosted AX, SSO and audit logs require Enterprise
Rank 5
Opik
Open-source LLM and agent observability, evaluation and testing platform from Comet, with a hosted cloud version.
comet.com8.6
out of 10
Best for: Teams that want open-source tracing plus pass/fail test suites, with the cheapest paid cloud plan on this list.
Opik is Comet's open-source platform for tracing, evaluating and monitoring LLM apps and agents. It traces multi-turn conversations and complex workflows, tracks tokens and cost, logs images, video and audio, and surfaces errors at the project level. It integrates with 40+ frameworks and providers, including LangChain, OpenAI, Google ADK, LangGraph and CrewAI, and supports OpenTelemetry for other languages. Its test suites let you write plain pass/fail assertions per item or for a whole suite, and it ships 30+ built-in metrics for relevance, context precision, task completion and hallucination. Annotation queues, an experiments dashboard, production guardrails and the OpikAssist debugging helper round it out.
The open-source build is free with no usage limits. Free Cloud covers up to 10 members, 25k spans per month and 60-day retention. Pro Cloud costs $19 per month for up to 50 members and 100k spans, with extra spans at $5 per 100k. That is the cheapest paid entry on this list.
Standout feature
Test suites turn plain-text pass/fail assertions into automated regression checks for agents.
Use cases
Small AI teams
Trace a CrewAI agent on Pro Cloud
Outcome: Tool calls, costs and errors are visible for $19 per month
QA engineers
Write pass/fail assertions for agent scenarios
Outcome: Each release runs a regression suite before it ships
Compliance teams
Apply guardrails for PII exposure in production
Outcome: Policy violations are blocked and logged
Key features
- Agent and conversation tracing with token and cost tracking and multi-media logging
- Integrations with 40+ AI frameworks, model providers and gateways, plus OpenTelemetry
- Test suites with item-level and suite-level pass/fail assertions
- 30+ built-in LLM-as-judge and heuristic evaluation metrics
- Annotation queues for subject matter experts and guardrails for PII and policy violations
Pros
- Open-source version shares the same codebase as the hosted product
- Pro Cloud costs $19 per month for up to 50 team members
- Free Cloud allows up to 10 team members
Cons
- Cloud plans keep data for 60 days
- Extra spans on Pro cost $5 per 100k
- Free Cloud is capped at 25k spans per month
Rank 6
W&B Weave
Weights & Biases toolkit for tracing, evaluating and monitoring production agents and LLM applications.
wandb.ai8.4
out of 10
Best for: ML and AI teams already using Weights & Biases for experiment tracking that want agent tracing in the same workspace.
Weave is the LLM and agent side of Weights & Biases, now sold under CoreWeave. It treats sessions, turns, steps, tools and sub-agents as first-class concepts, so a multi-agent run reads the way it executed. Built-in and custom signals classify interactions in production, and alerts go to Slack or trigger webhooks. The evaluation framework uses an imperative API with comparison views to catch regressions, and Guardrails includes scorers for toxicity, bias, PII, hallucination, coherence and relevance. A playground tests models against production traces, leaderboards aggregate eval results and an MCP server lets coding agents read live data and run evals.
The Free plan includes tracing, evals and scorers with 1 GB of Weave ingestion per month. Pro starts at $60 per month billed monthly with a 30-day trial, adds team access controls, CI/CD automations and alerts, and includes 1.5 GB of ingestion, with extra data at $0.10 per MB. Heavy tracing can get expensive.
Standout feature
Sessions, turns, steps and sub-agents are first-class objects in Weave traces.
Use cases
ML research teams
Trace fine-tuned models next to their training runs
Outcome: Model lineage and production behavior live in one workspace
Agent developers
Browse multi-agent sessions by turn and sub-agent
Outcome: Root causes buried in long runs are easier to find
Trust and safety teams
Score outputs with Guardrails for toxicity and PII
Outcome: Unsafe responses are flagged before users see them
Key features
- Agent-native tracing with sessions, turns, steps, tools and sub-agents
- Built-in and custom signals that classify agent interactions, with Slack and webhook alerts
- Imperative evaluation API with comparisons and visualizations
- Guardrails scorers for toxicity, bias, PII, hallucination and quality
- Playground, leaderboards and an MCP server for coding agents
Pros
- Same account covers model training experiments and LLM tracing
- 30-day free trial of Pro
- Free for academic research
Cons
- Pro includes only 1.5 GB of Weave ingestion per month
- Extra Weave ingestion costs $0.10 per MB
- Pro is meant for teams under 50 employees
Rank 7
Datadog LLM Observability
Agent and LLM monitoring inside Datadog that ties model calls to backend services, infrastructure and real user sessions.
datadoghq.com8.3
out of 10
Best for: Teams already running Datadog that want LLM traces next to their APM, logs and infrastructure data.
Datadog LLM Observability brings agent tracing into the Datadog platform. Each agent step is traced with span-level inputs, outputs and metadata, and latency, token usage and cost are monitored live. Evaluators, both built in and custom, check quality continuously and include hallucination and drift detection, and security scanning watches for prompt injection, data leaks and unsafe outputs. For development, it offers human annotation, versioned datasets from production traces, a playground and structured experiments with offline evaluators. Its main advantage is correlation: an agent problem can be followed into backend services, infrastructure metrics and real user sessions in the same tool.
Pricing counts only LLM spans, meaning calls to a model provider. The free tier covers up to 40K LLM spans a month with 15-day retention. Pro starts at $160 per month on annual pricing for 100K LLM spans, with extra spans at $3.50 per 10K. Retention add-ons extend traces to 90 days.
Standout feature
Agent traces link directly to APM, infrastructure and real user monitoring data in the same platform.
Use cases
SRE teams
Follow a slow agent response into backend traces
Outcome: A database bottleneck behind an AI feature is found quickly
Security teams
Scan LLM traffic for prompt injection
Outcome: Attack attempts are flagged alongside other security signals
Engineering managers
Track token cost per agent in dashboards
Outcome: AI spend is reported with the rest of the infrastructure bill
Key features
- Span-level tracing of prompts, retrieval, tool calls and decisions
- Real-time latency, token usage and cost monitoring across agents
- Out-of-the-box and custom evaluators, including hallucination and drift detection
- Security scanning for prompt injection, sensitive data exposure and unsafe outputs
- Correlation with backend services, infrastructure and real user sessions
Pros
- Billing counts only LLM spans, not tool, workflow or retrieval spans
- Full-stack root cause analysis with the rest of Datadog
- Sensitive Data Scanner included
Cons
- Retention is 15 days unless you buy add-ons
- Listed prices are annual; on-demand and month-to-month options are quoted separately
- Most valuable only if you also use other Datadog products
Rank 8
Helicone
Open-source AI gateway and LLM observability platform that logs every request routed through it.
helicone.ai8.2
out of 10
Best for: Teams that want logging, caching and provider fallbacks from a one-line gateway integration.
Helicone takes a gateway-first approach. You point the OpenAI SDK at its AI gateway, and every request to 100+ models from OpenAI, Anthropic, Google, Groq and others is logged automatically, with caching, custom rate limits and fallbacks if a provider fails. Gateway credits let you skip managing provider keys, and Helicone states a 0% markup on provider prices. On the observability side, it tracks sessions, user metrics, custom properties and cost, and Pro adds alerts, reports and HQL, its query language. Prompt management, a playground, scores and datasets cover basic iteration, though eval depth trails Braintrust and LangSmith.
Hobby is free with 10,000 requests, 1 GB of storage, 1 seat and 7-day retention. Pro costs $79 per month plus usage with unlimited seats and 1-month retention, and Team at $799 adds 5 organizations, SOC 2 and HIPAA compliance and 3-month retention. The site currently states Helicone is no longer part of Mintlify, so check for ownership updates.
Standout feature
A single base URL change both routes traffic to 100+ models and logs every request.
Use cases
Early-stage startups
Swap the OpenAI base URL for the Helicone gateway
Outcome: Every request is logged with cost in minutes
Teams on several providers
Set automatic fallbacks between models
Outcome: A provider outage does not take the product down
High-traffic apps
Cache repeated prompts at the gateway
Outcome: Duplicate requests stop costing tokens
Key features
- OpenAI-compatible AI gateway with access to 100+ models
- Automatic request logging with sessions, user metrics and custom properties
- Caching, custom rate limits and automatic provider fallbacks
- HQL query language, alerts and reports
- Prompts, playground, scores and datasets
Pros
- Setup takes a base URL change in the OpenAI SDK
- Gateway credits carry 0% markup on provider prices
- Unlimited seats on Pro
Cons
- Evaluation tools are lighter than eval-first platforms
- Hobby keeps data for 7 days with 1 seat
- The site notes that Helicone is no longer part of Mintlify, with details to come
Rank 9
Laminar
Open-source agent observability platform that captures LLM calls, tool calls and sub-agents and finds failures with its Signals model.
lmnr.ai8.1
out of 10
Best for: Teams running long agent workflows that want automatic failure detection across many runs.
Laminar is an open-source observability platform aimed squarely at agents. It captures LLM calls, tool calls, sub-agents, costs and tokens and presents each run as a readable transcript with a cost heatmap. Its distinguishing feature is Signals: a model tuned for trace analysis reads every run and flags failures you did not define in advance, such as repeated tool calls or answers built from unverified snippets. Similar failures are grouped into clusters, and Laminar can notify a Slack channel and answer questions about traces there. The platform also ingests OTLP traces and includes full-text search, a SQL editor, dashboards, evaluations, datasets, labeling queues and browser session recording.
Free includes 1 GB of data, 7 days of access, 1 project and 1 seat, without Signals. Starter costs $30 per month with 3 GB, 30 days and unlimited seats; Pro at $150 adds 10 GB and 6 months. Signals usage is billed per million tokens on top.
Standout feature
Signals analyze every agent run and cluster failure modes nobody had written a check for.
Use cases
Browser agent builders
Record browser sessions next to agent traces
Outcome: Each failed step is visible in context
Agent platform teams
Let Signals scan every production run
Outcome: Unexpected failure modes are clustered without writing rules
Small teams
Share one Starter workspace across the team
Outcome: Unlimited seats for $30 per month
Key features
- Automatic capture of LLM calls, tool calls, sub-agents, costs and tokens in a transcript view
- Signals that analyze agent runs to surface unexpected failure modes
- Signal clusters that group similar failures into patterns
- OTLP trace ingestion, full-text search, SQL editor and custom dashboards
- Evaluations, datasets, labeling queues and browser session recording
Pros
- Unlimited projects and seats from the $30 Starter plan
- Slack alerts for new failures and clusters
- On-premise option on Enterprise
Cons
- Signals analysis is billed separately by tokens
- Free plan keeps data for 7 days with 1 project and 1 seat
- Smaller ecosystem than LangSmith or Langfuse
Rank 10
HoneyHive
AI observability and evaluation platform for enterprises running agents in production, with SaaS, hybrid and self-hosted deployment.
honeyhive.ai8.0
out of 10
Best for: Enterprises in regulated industries that need audit-ready evidence and flexible hosting for agent monitoring.
HoneyHive targets companies running agents at scale, especially in regulated sectors. Its tracing is OpenTelemetry-native and vendor-agnostic, instrumenting 100+ models and frameworks automatically and covering in-house agents, coding agents and agents built on third-party platforms. Evaluation covers automated and human scoring, offline experiments against datasets built from production cases and annotation queues for domain experts. Alerts detect behavior drift, and prompt management handles versioning and deployment. Deployment options are a key selling point: multi-tenant SaaS, a hybrid setup where traces and evals stay in your cloud, or fully self-hosted on Kubernetes. HoneyHive is SOC 2 Type II audited, GDPR-compliant with EU residency and offers a HIPAA BAA.
The Developer plan is free with 10,000 events per month, up to 5 users, CI/CD integration and 30-day retention. Beyond that, only custom Enterprise pricing is published, which adds unlimited users, SSO, uptime SLAs and self-hosting. Smaller teams will hit the free limits with no self-serve step up.
Standout feature
Hybrid deployment keeps traces and evals in your own cloud while HoneyHive manages the rest.
Use cases
Banks and insurers
Run HoneyHive in hybrid mode
Outcome: Trace data stays in their cloud while the vendor runs the control plane
Domain experts
Define rubrics and review real agent outputs
Outcome: Expert judgment becomes reusable evaluation data
Risk and governance teams
Review traces, eval results and human reviews
Outcome: Audit-ready evidence for each agent release
Key features
- Distributed tracing that is OpenTelemetry-native and vendor-agnostic
- Automatic instrumentation for 100+ models and frameworks
- Automated and human evaluations, experiments and datasets from production cases
- Alerts for behavior drift, annotation queues and prompt management
- SaaS, hybrid and fully self-hosted deployment
Pros
- Free Developer plan includes CI/CD integration and up to 5 users
- SOC 2 Type II audited, GDPR-compliant and HIPAA BAA available
- Traces coding agents and third-party agent platforms too
Cons
- No self-serve paid plan between free and Enterprise
- Developer plan keeps data for 30 days
- Self-hosting and SSO need Enterprise
Frequently asked questions
+What is the best LLM observability tool in 2026?
Langfuse is our top pick because it combines tracing, prompt management and evals in an open-source product you can self-host, with a cloud Core plan at $29 per month for unlimited users. LangSmith and Braintrust are the strongest alternatives, for LangChain users and eval-heavy teams respectively.
+What is the best open-source LLM observability tool?
Langfuse, Arize Phoenix, Opik and Laminar all publish their code and can run on your own servers. Langfuse has the broadest feature set, Phoenix starts locally with one command, and Opik adds pass/fail test suites. Phoenix uses the ELv2 license, which limits offering it as a hosted service.
+How much do LLM observability tools cost?
Every tool here has a free tier. Paid plans start at $19 per month for Opik Pro Cloud, $29 for Langfuse Core, $30 for Laminar Starter, $39 per seat for LangSmith Plus, $50 for Arize AX Pro, $60 for W&B Pro, $79 for Helicone Pro, $160 for Datadog on annual pricing and $249 for Braintrust Pro. HoneyHive publishes only custom Enterprise pricing above its free plan.
+What is the difference between LLM observability and APM?
APM measures response times, errors and throughput across services. LLM observability records prompts, outputs, token counts, cost and tool calls for each model call and adds evaluation of output quality. Datadog offers both in one platform, while tools like Langfuse and LangSmith focus on the LLM layer.
+Do I need OpenTelemetry for LLM tracing?
No, but it helps. Langfuse, Arize Phoenix, Opik, LangSmith, HoneyHive and Laminar accept OpenTelemetry traces, which keeps your instrumentation portable between vendors. Native SDKs are usually quicker to set up for Python and TypeScript.
About this list
Updated November 9, 2026 by the Mojolin research desk. Mojolin is self-funded
Questions or corrections: editorial@mojolin.com
More best software lists