LLM observability tools record the prompts, tool calls, costs and scores of an AI application in production, and they are sold per seat, per event, per span, per record or per gigabyte.
This guide ranks twelve products on setup effort, the price once real traffic arrives, how data leaves and whether you can run it independently.
Vendors can fund visibility on this page. They cannot fund a better entry, since criticism stays in, and we are not paid when readers click through. How that works.
In short
What LLM observability tools software does
LLM observability software records traces, costs, latency and evaluation scores for language-model applications and agents, either in the vendor's cloud or on infrastructure that you operate.
Setup effort, true cost, data portability, exit options and buyer fit are the five tests, applied in that order to every LLM observability tool here. One rule sits on top of those tests: places 1 to 3 go to challengers, vendors outside the best-known brands in a category, wherever the category has them; the established names follow from place 4 in their ranked order. Why those five, and why there is no score out of ten, is on the how we work page.
Euro-priced agent tracing and simulation from Amsterdam, with Apache-2.0 code
Ranked #1 of 12 in Best LLM Observability Tools in 2026.
Free tierPublished pricingEurope
LangWatch bundles tracing, evaluations, prompt management and agent simulations, and the pricing page lists simulated-user testing as available in every package, the free one included.
The repository carries Apache-2.0, and Enterprise adds hybrid, self-hosted or on-premises routes. Billing has two meters, core seats and events, so a small review group with heavy traffic pays mostly for events. ISO 27001 reports and a custom DPA are part of Enterprise only.
What stands out
Apache-2.0
Agent simulations
Per core seat
Where it costs you
Data access on the Developer plan lasts 14 days and Growth retention is 30 days, so older traces cost extra per gigabyte.
Seats are counted as core seats, so the number of people who edit and review sets the invoice before any events are added.
Right for
Teams that want agent simulations and euro invoicing next to plain trace logging
Wrong for
Buyers who need long history without paying for storage
NetherlandsFree Developer plan, then a monthly fee per core seat plus event volume; prices published
Agent tracing sold by gigabytes of stored data instead of by seat
Ranked #2 of 12 in Best LLM Observability Tools in 2026.
Free tierSelf-hostablePublished pricingNorth America
Laminar sells a single product, observability for agents, and sets cost by how many gigabytes of trace data you keep. A large review group stays cheap because paid tiers include unlimited seats, while an agent with long contexts fills the allowance faster.
Pro keeps six months of data and Starter thirty days. Enterprise lists on-premise deployment and custom limits, with no amount on the page. The company is LMNR AI, Inc., governed by Delaware law.
What stands out
Apache-2.0
Per gigabyte
Signals analysis
Where it costs you
The free plan stores 1 GB with no overage and keeps data 7 days, a short window for chasing a regression that surfaced late.
Signals analysis is billed per token on paid plans, which puts a second line on the invoice beside the plan fee.
Right for
Engineering groups with many reviewers who would rather pay by stored data
Wrong for
Organisations that need one fixed annual figure with no meters
United StatesFree plan, then monthly tiers by data volume with per-gigabyte overage; prices published
Tracing and prompt management billed in units, now owned by ClickHouse
Ranked #3 of 12 in Best LLM Observability Tools in 2026.
Free tierSelf-hostablePublished pricingNorth America
Langfuse puts tracing, prompt management, evaluation and dashboards in one product, with SDKs for Python and JavaScript and OpenTelemetry for other languages. Cloud data access grows from 30 days on Hobby to 90 on Core and three years on Pro.
SSO sits behind a monthly add-on on Pro, while audit logs and SCIM start at Enterprise. Unit volume drives the cost, so an agent emitting many observations per request multiplies the bill.
What stands out
MIT core
Unit-based billing
Self-hostable
Where it costs you
Hobby keeps data 30 days and allows two users, so a working team moves to a paid plan quickly.
Self-hosted installs need a licence key for add-on features, and the Docker Compose route is described as a single VM without high availability or backups.
Right for
Teams that want the option to move from hosted to their own infrastructure
Wrong for
Buyers who want a flat price with no per-unit overage
United StatesFree Hobby plan; monthly tiers with per-unit overage; prices published
OpenTelemetry observability for agents and services, billed per telemetry record
Ranked #4 of 12 in Best LLM Observability Tools in 2026.
Free tierSelf-hostablePublished pricingNorth America
Logfire accepts anything instrumented with OpenTelemetry, so Pydantic AI, LangChain and OpenAI SDK traces sit beside database and host data. Beyond five seats the record meter drives cost, at a stated rate per extra million.
An AI gateway lets you bring your own provider keys without markup or use built-in providers with a 5% markup. Enterprise adds SSO, SCIM, SLOs and the choice of cloud, dedicated or self-hosted.
What stands out
OpenTelemetry
Per record
Enterprise self-hosting
Where it costs you
The free plan has one seat and three projects, and ingestion pauses at the cap instead of overflowing.
Logs, metrics and spans count together as records, so a noisy service can use up the allowance meant for model calls.
Right for
Python teams already instrumenting services with OpenTelemetry who want model calls on that timeline
Wrong for
Groups wanting a tool built only for prompt iteration
United StatesFree plan with a hard cap; Team per month plus per-record overage; prices published
Apache-2.0 tracing and evaluation from Comet, free to run yourself
Ranked #5 of 12 in Best LLM Observability Tools in 2026.
Free tierSelf-hostablePublished pricingNorth America
The pricing page shows Opik in three forms: the open code, a free cloud plan for up to ten members, and Pro for up to fifty. Included are agent tracing, test suites with assertions and an agent playground.
Span volume sets the tier, so an agent with many tool calls reaches 25,000 spans sooner than a chatbot. Enterprise lists SOC 2, ISO 27001, ISO 9001, HIPAA and GDPR compliance.
What stands out
Apache-2.0
Span-based plans
Self-hostable
Where it costs you
Both cloud plans keep data 60 days unless Pro retention is customised, so long audits need a paid change.
SSO, service accounts and dedicated support are limited to the Enterprise plan, which has no listed price.
Right for
Teams that prototype on the hosted plan and want a self-hosted exit later
Wrong for
Organisations that need single sign-on without negotiating a contract
United StatesFree cloud plan; Pro per month; Enterprise quoted; prices published
Agent tracing and online evaluation with an event-capped Developer plan
Ranked #6 of 12 in Best LLM Observability Tools in 2026.
Free tierSelf-hostablePublished pricingNorth America
HoneyHive's product list covers tracing, online evaluations on live traffic, monitoring, alerts, offline experiments, datasets, annotation queues and prompt management. The site features Australia's largest bank as a customer.
Data stays 30 days on Developer, the plan also includes prompt versioning and CI/CD integration, and retention above that is set by contract. HoneyHive Inc. is a Delaware corporation whose terms follow New York law, a point procurement will read.
What stands out
Event-based
Online evaluations
Optional self-hosting
Where it costs you
The Developer plan allows 10,000 events a month, which a busy agent can exhaust quickly.
No mid-tier sits between Developer and Enterprise, so growth means a sales contract.
Right for
Teams testing agent quality on live traffic before signing a vendor contract
Wrong for
Buyers wanting a published price for production volume
United StatesFree Developer plan; Enterprise quoted per organisation
Source-available tracing you run yourself, with the Arize AX cloud beside it
Ranked #7 of 12 in Best LLM Observability Tools in 2026.
Free tierSelf-hostablePublished pricingNorth America
Phoenix and Arize AX are two products from one company. Phoenix is the code you deploy through Docker or Kubernetes, and its documentation covers Aurora and Azure PostgreSQL backends.
AX is the managed platform, with unlimited users and evals on every plan, while spans, ingestion and retention differ by tier. The free tier stops at one gigabyte of ingestion a month, so plan any migration before volume grows.
What stands out
Elastic License 2.0
Self-hosted
Arize AX cloud
Where it costs you
The Elastic License 2.0 prohibits hosting Phoenix as a managed service for third parties, which rules out reselling it.
AX Pro includes 50,000 spans, 10 GB and 30 days of retention a month, small for production traffic.
Right for
Evaluation-focused engineers who want to inspect traces locally before buying a cloud plan
Wrong for
Resellers who want to embed it in a managed offering
United StatesFree to run yourself; Arize AX free and Pro cloud plans; Enterprise quoted; prices published
Evaluation-first platform billed by processed data and scored outputs
Ranked #8 of 12 in Best LLM Observability Tools in 2026.
Free tierSelf-hostablePublished pricingNorth America
Braintrust treats tracing as the input to evaluation, and the pricing page lists a built-in agent, Loop, that can run evaluations, write test cases and iterate on prompts. Qualifying startups get six to twelve months of Pro free.
Starter includes ten dollars of model credits and Pro one hundred, which cover built-in AI features, so teams that bring their own keys draw less. Braintrust Data, Inc. is based in San Francisco.
What stands out
Processed data
Scores metered
Unlimited users
Where it costs you
Retention is 14 days on Starter and 30 days on Pro, and extra Pro retention is billed per gigabyte each month.
Model-judge and code scorers are metered per thousand scored outputs, so heavy online scoring raises the bill.
Right for
Teams whose main job is running evaluations against changing prompts
Wrong for
Anyone who needs to forecast spend from seat count alone
United StatesFree Starter plan; Pro per month plus data and score usage; prices published
AI gateway with request logs built in, priced by recorded requests
Ranked #9 of 12 in Best LLM Observability Tools in 2026.
Free tierPublished pricingNorth America
Portkey treats observability as one module among several: a universal API with automatic fallbacks, load balancing and retries, prompt management, guardrails and an MCP gateway. Governance on Enterprise covers role-based access, SSO, budgets and rate limits.
The gateway code is MIT-licensed on GitHub. Observability, key management, prompt management and routing are listed as supported on all plans. Anyone who wants only trace storage should compare it with tools that ingest OpenTelemetry directly.
What stands out
AI gateway
MIT gateway
Request-based
Where it costs you
Developer logs last three days and the page itself calls that plan unsuitable for production workloads.
Past 3 million requests a month the page lists no price, only a custom Enterprise contract.
Right for
Teams routing several model providers through one gateway who want logs included
Wrong for
Groups wanting a dedicated evaluation and dataset workbench
United StatesFree Developer plan; Production per month plus request overage; Enterprise quoted; published
Per-user LLM logging with prompt management, sold as a paid Team plan
Ranked #10 of 12 in Best LLM Observability Tools in 2026.
Self-hostablePublished pricingElsewhere
Lunary's Team plan lists prompt management, human reviews, custom dashboards, smart views, topics, CSV and JSONL exports and full API access. Events beyond the 50,000 included are billed per 50,000 each month, and AI playground queries past 300 cost five cents each.
History lasts a year on Team and is set by contract on Enterprise, which adds data-warehouse connectors, an audit trail, white labelling and a technical account manager.
What stands out
Per user
50,000 events
Enterprise self-hosting
Where it costs you
New accounts have no self-serve entry level, since payment is required before workspace access.
The Team plan caps seats at ten, so a larger group has to negotiate Enterprise.
Right for
Product teams of up to ten people who want logs and human review together
Wrong for
Hobby projects or evaluations that need a zero-cost start
Not stated by the vendorPer user per month; extra events billed per block; Enterprise quoted; prices published
LangChain's tracing service with seat fees and pay-as-you-go trace billing
Ranked #11 of 12 in Best LLM Observability Tools in 2026.
Free tierSelf-hostablePublished pricingNorth America
LangSmith combines a per-seat fee with metered traces: Developer gives one seat and 5,000 base traces a month, Plus charges per seat with 10,000 base traces, and usage beyond that is billed in LangChain Standard Units.
Self-hosted and hybrid deployment are Enterprise options. The vendor ships a usage calculator because trace volume and deployments move the bill, so a headline seat price predicts little.
What stands out
Per seat
Pay-as-you-go traces
Self-hosted Enterprise
United StatesFree Developer seat; Plus per seat plus pay-as-you-go traces; Enterprise quoted; published
LLM span metering inside an existing Datadog account, now named Agent Observability
Ranked #12 of 12 in Best LLM Observability Tools in 2026.
Pricing on requestNorth America
Datadog's documentation now calls the product Agent Observability and bills it on the number of LLM spans ingested. One agent workflow can produce several spans, so spend rises with workflow depth.
It sits beside Datadog's infrastructure and application monitoring, which helps where those teams already pay Datadog, but a team needing only LLM traces buys into a much larger platform. The docs list an EU site among the regions.
What stands out
Per LLM span
Datadog platform
OpenTelemetry
United StatesMetered per LLM span ingested, added to a Datadog contract
LLM observability software records traces, costs, latency and evaluation scores for language-model applications and agents, either in the vendor's cloud or on infrastructure that you operate. Differences that change the outcome seldom sit in the features. Here is the sequence we would follow.
01
Decide whether you need a published price
The split is 11 published against 1 quoted. If you want to compare numbers first, start with LangWatch, Laminar, Langfuse, Pydantic Logfire, Opik, HoneyHive, Arize Phoenix, Braintrust, Portkey, Lunary, LangSmith.
02
Decide how much the jurisdiction matters
Where a company is established decides which courts and regulators can reach it. Here that means 3 countries: North America 10, Europe 1, Elsewhere 1.
How LLM observability pricing works: events, spans, units, records or gigabytes
Each vendor counts something different, and the unit decides which agent design is expensive. LangWatch bills core seats plus events, Laminar bills stored gigabytes with unlimited seats, Langfuse bills units, Pydantic Logfire bills telemetry records shared with logs and metrics, and Datadog bills LLM spans inside an existing contract.
An agent that makes ten model calls per user request produces ten spans but may produce a single trace, so the same traffic costs different amounts under each meter. Before comparing price pages, measure one real week of your own traffic: requests, calls per request and average payload size, then convert that profile into each vendor's unit and ask for the result in writing.
Count how many spans, observations or events one user request creates in your application
Check whether logs and metrics share the same allowance as model traces
Ask what happens at the cap: a hard stop, an overage rate or a forced plan change
Retention windows: from 3 days to 3 years
Free plans in this market keep data for a short time, and the gap between plans is wide. Portkey's Developer plan keeps logs three days, Laminar's free plan seven, LangWatch's Developer plan fourteen, Braintrust Starter fourteen and Arize AX free fifteen. Langfuse moves from 30 days on Hobby to 90 on Core and three years on Pro, while Opik keeps 60 days on both cloud plans.
A regression found after a quarterly release review can fall outside a two-week window, and extra retention is often billed per gigabyte rather than included. Decide how far back you must look for audits, customer complaints and evaluation baselines, then price that window on each shortlisted plan, not the entry plan.
Write down how many days of traces an incident or audit would need
Check whether longer retention is a plan upgrade or a per-gigabyte add-on
Confirm that exported data keeps spans, scores and metadata, not only raw text
Running it yourself: Opik, Langfuse, Arize Phoenix and the licence behind them
Self-hosting is a common reason to shortlist these tools, but the licences differ. Opik is Apache-2.0 for the whole platform, LangWatch and Laminar are Apache-2.0, and Langfuse is MIT with an enterprise directory and licence-key add-ons. Arize Phoenix uses the Elastic License 2.0, which bars offering it as a managed service.
Other vendors reserve self-hosting for contracts: Pydantic Logfire, Lunary, HoneyHive and LangSmith list it on Enterprise plans. Running a tool yourself moves cost to databases, upgrades and on-call time, so a free licence is not a free deployment. Read the repository licence file, not the marketing page, before you promise a customer that the data never leaves your network.
Read the licence text in the repository and note any directory under a different licence
List which features need a licence key or an Enterprise contract when self-hosted
Estimate the database and upgrade effort your team would carry
Tracing versus evaluation: Braintrust, HoneyHive, LangSmith and Portkey compared
Some products centre on storing traces, others on judging outputs, and the invoice follows that focus. Braintrust meters scored outputs separately, so running a model judge on every trace adds a line of its own. HoneyHive runs online evaluations on live traffic within an event allowance. LangSmith charges per seat and per trace, while Portkey logs requests that pass through its gateway and prices them per request.
Decide whether the first job is debugging production runs, scoring prompt changes before release or controlling provider spend, because buying an evaluation workbench to read logs, or a gateway to run experiments, leaves half the product unused. Pilot one workflow in two tools before signing for a year.
Name the first job: debugging, pre-release evaluation or cost control
Check whether scoring is metered separately from trace storage
Run the same agent through two shortlisted tools for a week and compare the effort
What goes wrong most often when buying LLM observability tools software
Choosing on the entry price alone, when free plans keep data for as little as three days and cap seats or events well below production use.
Treating a free licence as a free deployment, then discovering that the database, upgrades and on-call work for a self-hosted stack cost more than the cloud plan.
Ignoring seat caps and role limits, such as Lunary's ten seats or Opik's member limits, until a second team asks for access.
Sending prompts and customer data to a hosted tracer without checking what the vendor stores, for how long and whether masking or zero retention is available.
07
Frequently asked questions
7 answers
What is the best LLM observability tools in 2026?
Our ranking of 12 puts LangWatch first. LangWatch is run by Reasoning Engine B.V. in Amsterdam and prices in euros: a monthly fee per core seat, plus events, of which 200,000 are included before a per-100,000 rate applies.
The code is Apache-2.0 and can be run yourself. Retention is the catch, since data access lasts 14 days on the free Developer plan and 30 days on Growth, with extra storage billed per gigabyte.
Which LLM observability tools publish their pricing?
11 of the 12, with the pricing model each one publishes:
LangWatch: Free Developer plan, then a monthly fee per core seat plus event volume; prices published.
Laminar: Free plan, then monthly tiers by data volume with per-gigabyte overage; prices published.
Pydantic Logfire: Free plan with a hard cap; Team per month plus per-record overage; prices published.
Opik: Free cloud plan; Pro per month; Enterprise quoted; prices published.
HoneyHive: Free Developer plan; Enterprise quoted per organisation.
Arize Phoenix: Free to run yourself; Arize AX free and Pro cloud plans; Enterprise quoted; prices published.
Braintrust: Free Starter plan; Pro per month plus data and score usage; prices published.
Portkey: Free Developer plan; Production per month plus request overage; Enterprise quoted; published.
Lunary: Per user per month; extra events billed per block; Enterprise quoted; prices published.
LangSmith: Free Developer seat; Plus per seat plus pay-as-you-go traces; Enterprise quoted; published.
The other 1 quote per organisation.
Is there a free LLM observability tool?
LangWatch, Laminar, Langfuse, Pydantic Logfire, Opik, HoneyHive, Arize Phoenix, Braintrust, Portkey, LangSmith can be used for free, either as a free plan or a self-hosted edition.
Where are these LLM observability tool vendors established?
Vendor countries on this page, North America 10, Europe 1, Elsewhere 1:
LangWatch: Netherlands.
Laminar: United States.
Langfuse: United States.
Pydantic Logfire: United States.
Opik: United States.
HoneyHive: United States.
Arize Phoenix: United States.
Braintrust: United States.
Portkey: United States.
Lunary: Not stated by the vendor.
LangSmith: United States.
Datadog: United States.
Which LLM observability tools can you host yourself?
Laminar, Langfuse, Pydantic Logfire, Opik, HoneyHive, Arize Phoenix, Braintrust, Lunary, LangSmith. The other 3 are hosted by the vendor only.
What should you use instead of LangWatch?
Laminar and Langfuse are the next two on this page. Laminar is for Engineering groups with many reviewers who would rather pay by stored data; Langfuse is for teams that want the option to move from hosted to their own infrastructure.
Who should not buy LangWatch?
Buyers who need long history without paying for storage. Data access on the Developer plan lasts 14 days and Growth retention is 30 days, so older traces cost extra per gigabyte..
This list covers 12 LLM observability tools products. A product we left out can be put to us by its maker, with a short note on its purpose and its customers; we decide on inclusion ourselves. The rules are under software advice.
