Published by Winding AI, Inc., the company that makes Griffin. No vendor paid to be listed.Prices read November 19, 2026 · How this site works
LLM Observability Tools10 tools compared · priced per 100k · read Nov 19See the tools
Prices read on November 19, 2026

The best AI observability tools, compared on price

For most teams the answer is Langfuse or Opik. Of the ten AI observability and LLM monitoring tools below, they and Arize are the only ones that trace each agent step, run evaluations, manage prompts, price each call and self-host free, and their paid plans start at $29 and $19 a month.

→LangSmith suits LangChain teams, Datadog suits teams already on Datadog, PostHog suits teams that want it beside product analytics. Griffin suits teams that only need what each call costs per feature, plus a test of cheaper models.

  • 14 vendors' pricing pages read on one day
  • Prices put per 100,000 spans where they can be
  • No vendor paid to be listed

Updated November 20, 2026. Also called LLM monitoring tools.

  • $5 to $500what 100,000 more records cost (Opik spans to LangSmith traces)
  • 5 of 14tools bought by another company in 2026
  • 11 of 14accept OpenTelemetry (OTLP) traces directly
  • 24 h to 60 dayshow long free plans keep full traces
What it is

What AI observability tools do

An AI observability tool for LLM apps records every model call and every agent step around it: prompt, response, tokens, cost, latency, errors. Calls nest into traces, so one bad answer opens onto the tool call that caused it. Most tools add evaluations, prompt versions and alerts.

What differs is the unit you pay for: spans (Opik, Arize), units (Langfuse), whole traces (LangSmith), events (PostHog), LLM spans only (Datadog) or gigabytes (Laminar, Braintrust, Weave). The same month can cost $64 or $489.

  • traceOne request or agent run from start to finish, holding its spans.
  • spanOne step inside a trace: an LLM call, a tool call, a retrieval.
  • unit (Langfuse)A trace, an observation or a score; each one counts.
  • OTLPThe OpenTelemetry wire format; a tool that accepts it can take traces from any OTel SDK.
  • GenAI conventionsOpenTelemetry's names for LLM spans and tokens; no cost field.
Compare

AI observability tools side by side

Prices from each vendor's own pricing page on November 19, 2026, in US dollars, before discounts. The data page adds Helicone, Lunary, Traceloop and Portkey, with sources and a CSV.

Ten tools, in this page's order Free plan keeps = retention on the free plan
ToolFree planFirst paid planPrice of 100k moreFree plan keepsSelf-hostOTLP in
Opik25k spans a month$19/mo flat, 100k spans$5 per 100k spans60 daysYes, Apache-2.0Yes
Langfuse50k units a month$29/mo, 100k units$8 per 100k units30 daysYes, MIT coreYes
Arize25k spans + 1 GB$50/mo, 50k spansNot published15 daysPhoenix, ELv2Yes
PostHog100k events a monthUsage only, no plan fee$26.25 for the next 100k; $8 per 100k by 1M1 year (text 30 days)UnsupportedYes
Laminar1 GB a month$30/mo, 3 GB$2 per GB7 daysYes, Apache-2.0Yes
LangSmith5k traces, 1 seat$39 a seat, 10k traces$500 per 100k traces14 daysEnterprise add-onYes
W&B Weave1 GB a monthFrom $60/mo, 1.5 GB$100 per GBNot statedLicensed onlyYes
Datadog40k LLM spans$160/mo yearly, 100k$35 per 100k LLM spans15 daysNoYes
Braintrust1 GB + 10k scores$249/mo, 5 GB$3 per GB14 daysBYOC on EnterpriseYes
GriffinWatch: every callRoute: 5% of routed spend$0 to watchBodies 7 daysNo (MIT SDK)No

Swipe the table sideways to see every column.

The tools

Ten AI observability tools, one card each

Order: by how many of five jobs a tool does on a plan you can start without sales (traces of each step, evaluations, prompt management, cost per call, a free self-hosted build); ties go to the cheaper paid plan.

1

Opik

Best for a low flat price
Paid from$19/mo

Comet's open-source tracing and evaluation platform. Free: 25,000 spans a month, kept 60 days. Pro: $19 a month flat for 100,000 spans, then $5 per 100,000.

Good

  • Lowest paid plan and cheapest extra spans of the ten
  • Apache-2.0; OTLP, agent graphs, evaluations and a prompt library

Watch out

  • Self-hosted Opik has no user management
  • 400-day retention costs $29 per 100,000 spans more
2

Langfuse

Best for open-source self-hosting
Paid from$29/mo

Open-source tracing, evaluations and prompts, owned by ClickHouse since January 2026. Hobby: free, 50,000 units a month, 30 days. Core: $29 a month for 100,000 units, then $8 per 100,000.

Good

  • MIT core: self-host all core features free, SSO included
  • 37,651 GitHub stars; OTLP, agent graphs, MCP tracing

Watch out

  • Each trace, observation and score is a unit: a 10-step run uses 11+
  • Audit logs and retention controls need the paid licence
3

Arize (Phoenix and AX)

Best for free self-hosting, no feature gates
Paid from$50/mo

Phoenix is free to self-host; AX is the hosted platform. AX Free: 25,000 spans and 1 GB, kept 15 days. AX Pro: $50 a month for 50,000 spans, kept 30 days. Owned by Dynatrace since 1 October 2026.

Good

  • Phoenix self-hosting has “no feature gates”
  • OTLP in; cost rolled up to the trace and the project

Watch out

  • Phoenix is Elastic License 2.0: source-available, not open source
  • No published price for spans past the AX plan
4

PostHog AI Observability

Best next to product analytics
Pay as you go$0 + usage

LLM tracing inside PostHog's product analytics. The first 100,000 events a month are free; the next 50,000 cost $0.00035 each, then the price steps down to $0.000175, $0.0001 and $0.00008 an event by 1M. No seat charges.

Good

  • Cost per call, evaluations, prompts and sessions beside product analytics
  • Metadata kept 1 year on the free plan; OTLP over HTTP

Watch out

  • Prompt and response text kept 30 days on every plan
  • Self-hosting is “officially unsupported”
5

Laminar

Best for long retention on a small plan
Paid from$30/mo

Open-source agent tracing billed by data size. Free: 1 GB a month, 7 days, 1 seat. Pro: $30 a month for 3 GB, 3 months of data and unlimited seats, then $2 a GB.

Good

  • Apache-2.0; traces LLM calls, tools and control flow
  • Dollar cost on every LLM span; OTLP over gRPC and HTTP

Watch out

  • No prompt management
  • Billed by the GB, so long prompts cost more
6

LangSmith

Best for LangChain and LangGraph teams
Paid from$39/seat

LangChain's tracing and evaluation platform. Developer: free, 1 seat, 5,000 base traces a month. Plus: $39 a seat a month with 10,000 base traces in total, then $0.005 a trace. Base traces are kept 14 days.

Good

  • Every agent step is a run; turns link up as threads
  • OTLP, LLM-as-judge evals, prompt environments

Watch out

  • $500 per 100,000 extra traces, the dearest volume here
  • Self-hosting is an Enterprise add-on only
7

W&B Weave

Best if you already use Weights & Biases
Paid from$60/mo

Weights & Biases' tracing and evaluation toolkit, now sold as part of CoreWeave Forge. Free: 1 GB of data a month. Pro: from $60 a month with 1.5 GB, then $0.10 a MB.

Good

  • Agent timelines with nested tool calls; MCP tracing
  • OTLP endpoint; SOC 2 Type II on every plan

Watch out

  • $0.10 a MB is $100 a GB, the highest data price here
  • No retention stated; self-hosting only on a paid licence
8

Datadog LLM Observability

Best if you already run Datadog
Billed yearly$160/mo

Datadog's LLM product, now called Agent Observability. Free: 40,000 LLM spans a month, kept 15 days. Pro: from $160 a month billed yearly for 100,000 LLM spans, then $3.50 per 10,000.

Good

  • Only LLM spans are billed, not tool or retrieval spans
  • Beside your APM, logs and alerts; MCP client tracing

Watch out

  • $200 a month month-to-month, $240 on demand
  • No self-hosting; 15-day retention unless you pay more
9

Braintrust

Best for evaluation-led teams
Paid from$249/mo

An evaluation-first platform with tracing. Starter: free, 1 GB of data, 10,000 scores, 14 days. Pro: $249 a month for 5 GB and 50,000 scores, then $3 a GB.

Good

  • Cost rolls up from child spans to the parent
  • Unlimited users on every plan; MCP server tracing

Watch out

  • The highest entry price of the ten; scores billed on top
  • No self-hosting: BYOC on Enterprise keeps a SaaS control plane
10

Griffin

Best for cost per use case
Watch plan$0

A TypeScript fetch wrapper from Winding AI that records what each AI call cost, grouped by use case, after the call finishes. It isn't a tracing tool: one record per call, no spans. It also replays real calls on cheaper models with a judge (replay grading).

Price: Watch $0. Test: $10 of replays free, then $20 packs. Route: 5% of routed spend after the first $50.

Good

  • Replays real calls on cheaper models, grades each answer, routes only to a model that passed, and falls back to the original model if a routed call fails
  • Cost, tokens, latency and errors per call, by use case, recorded outside the request path

Watch out

  • One of the five jobs: no traces, eval suite or prompt management
  • Hosted only (MIT SDK); the wrapper is TypeScript

Not in the ten, but on the data page: Helicone (maintenance mode since March 2026), Lunary (no free plan for new sign-ups), Traceloop (paid plan is Enterprise only; joining ServiceNow) and Portkey (its pricing page now leads to a Palo Alto Networks sales form).

Cost at volume

What one busy month costs on each tool

A sample month: 100,000 agent runs, each one trace of 10 spans (4 LLM calls, 5 tool calls, 1 agent span). That is 1,000,000 spans and 400,000 LLM calls. List prices, one seat, monthly billing unless noted.

  • Opik$64$19 + 9 × $5 for 1M spans
  • PostHog$1001M events on the stepped price
  • Langfuse$1081.1M units: traces plus observations
  • Datadog$265400k LLM spans, billed yearly
  • LangSmith$489100k traces, one seat

Griffin ($0 to watch) records calls, not spans, so it isn't in the bars. Arize AX publishes no price past its plan. Laminar, Braintrust and W&B Weave bill by the GB: at about 4 GB a month (our estimate) they come to about $30, $249 and $310. The arithmetic is on the data page.

Choose

How to choose an AI observability tool

  1. Find your billing unit.Count a month of spans, traces and gigabytes, then price it on two or three tools.
  2. Decide whether you must self-host.Langfuse, Opik and Laminar run on your servers free; Phoenix too, under the Elastic License. LangSmith and Weave need paid licences; Datadog and Braintrust don't self-host.
  3. Check retention.Free plans keep full traces from 24 hours (Traceloop) to 60 days (Opik).
  4. Instrument with OpenTelemetry.11 of 14 tools take OTLP, so an OTel SDK keeps you free to switch.
  5. If the question is only cost, you may not need traces.A cost tool such as Griffin, or the cost views above, shows what each feature spends.
FAQ

Questions people ask about LLM observability

What is LLM observability?

Recording each language-model call and each step of an AI agent (prompt, response, tokens, cost, latency, errors) as traces you can search, score and alert on. Most tools add evaluations and prompt management.

Which AI observability tools are open source?

Langfuse (MIT core, 37,651 GitHub stars) and Opik (Apache-2.0, 23,458 stars) self-host traces, evaluations and prompt management under an open licence. Laminar and Helicone are Apache-2.0 too. Arize Phoenix is free to self-host under the Elastic License 2.0, which is source-available.

How much does LangSmith cost?

Developer is free for 1 seat and 5,000 base traces a month. Plus is $39 a seat a month with 10,000 base traces in total; each extra trace is 0.005 LSU (1 LSU = $1), so 100,000 more cost $500. Base traces are kept 14 days. Enterprise is quote-only.

Is Langfuse free?

Yes: self-hosting the MIT core is free, and Langfuse Cloud's Hobby plan is free for 50,000 units a month (traces, observations and scores) with 30 days of data. Core is $29 a month for 100,000 units, then $8 per 100,000.

Do AI observability tools work with OpenTelemetry?

11 of the 14 tools on the data page accept OTLP traces. OpenTelemetry's GenAI conventions name spans and token counts but have no cost attribute (and are still marked Development), so each tool prices tokens itself.

What happened to Helicone?

Mintlify announced on 3 March 2026 that it was buying Helicone and keeping it in maintenance mode. Since 12 November helicone.ai says “Helicone is now part of Experiential Labs. Details to come soon.” It is on the data page.

Is Griffin an AI observability tool?

Not in the tracing sense. Griffin records one event per AI call (model, tokens, cost, latency, errors) grouped by use case, with no spans, and adds replay grading and routing to cheaper models. It is made by Winding AI, Inc., which publishes this site.

About this page and how it was checked

Published by Winding AI, Inc., the company that makes Griffin. Winding AI makes money when teams pay for Griffin and from nothing else here: no vendor paid to be listed, and no link earns a commission. Griffin gets the same card, row and order rule as every tool, and comes last because it does one of the five jobs.

How this page was checked: on November 19, 2026 we read each vendor's pricing page, plan limits, self-hosting and OpenTelemetry docs, trust page and licence file, and counted GitHub stars through the GitHub API. Per-100,000 figures and the sample month are our arithmetic on list prices. Every source is in the source list. Corrections: support@winding.ai.