What are the best LLM observability tools in 2026?

Langfuse is where most teams start, and many stay. It calls itself an open source AI engineering platform, and it covers the three jobs you need once an LLM feature is live: tracing every call and agent step, scoring outputs (LLM-as-a-judge, code checks, human labels), and versioning prompts so you know which one produced which answer. It is built on OpenTelemetry, so any OTel-instrumented app can send it traces.

The license is the other reason. The repository is MIT licensed, except for its ee folders, and the vendor says you can self-host it for free. On the cloud, Hobby is free with 50k units a month, 30 days of data access and 2 users; Core is $29 a month with 100k units, 90 days of data access and unlimited users; Pro is $199 a month with 3 years of data access. Extra usage costs $8 per 100k units up to 1M, then $7 per 100k. Count units carefully: a trace, each step inside it and each score are all billable units.

In a 2025 developer thread comparing it with Phoenix, an engineer praised its out-of-the-box Python decorator for basic tracing. ClickHouse acquired Langfuse in 2026; the team says its roadmap stays the same and it remains committed to open source. The trade-off is that a self-hosted install is yours to run and upgrade.

Arize comes in two parts. Phoenix is its open-source, self-hosted AI observability platform for experimentation, evaluation and troubleshooting, built on top of OpenTelemetry and released under the Elastic License 2.0. AX is the managed cloud: Free with 25k spans and 1 GB of ingestion a month, 15 days of retention and unlimited users; Pro at $50 a month with 50k spans, 10 GB and 30 days.

Its strength is evaluation: templates for LLM-as-a-judge, datasets and experiments that data scientists find natural. In 2025 developer threads, a startup engineer said Phoenix was the only solution that worked out of the box with OTel; another found it clunky next to Langfuse, and two complained that spans sent through plain OTel showed up as "unknown".

Note the license: Elastic License 2.0 lets you self-host for your own use, but it is not an OSI open-source license like MIT or Apache.

LangSmith is LangChain's tracing and evaluation product, and on a LangChain or LangGraph app it is the shortest path to readable traces: a few environment variables and every chain, tool call and graph step shows up nested. Datasets, experiments and online evaluators sit next to the traces, so a bad production answer becomes a test case in a few clicks.

Pricing is per seat. Developer is free for one seat with 5k base traces a month; Plus is $39 per seat per month with 10k traces included, billed monthly. A team of five on Plus pays $195 a month before extra traces, and each trace beyond the included ones costs 0.005 LSU, at $1 per LSU: $5 per 1,000 traces. Seats added mid-month are pro-rated, but a removed seat gets no credit. VC-backed startups can get up to $10,000 in credits. It also accepts traces from apps that don't use LangChain, but its edge is clearest inside that ecosystem.

Read our LangChain review for the framework side, and our LangChain alternatives if you're moving away from it.

Opik is Comet's open-source LLM evaluation and tracing platform. You log traces and spans, run LLM-as-a-judge metrics, keep datasets and compare experiments, with integrations for the common frameworks. Comet says its core observability and evaluation features are free in the open-source code, under the Apache 2.0 license.

Its pricing is the easiest to say yes to: Free Cloud gives 25k spans a month for up to 10 team members with 60 days of retention; Pro Cloud is $19 a month for 100k spans and up to 50 members, then $5 per 100k extra spans. The Comet name comes from MLOps experiment tracking, so ML teams will find the concepts familiar.

The community around it is smaller than Langfuse's, so expect fewer answered questions when you get stuck.

Braintrust puts evaluation at the center. You build datasets, write scorers, run experiments on each prompt or model change and compare them side by side; production logs feed back into the datasets. Its UI is one of the most polished in this list, which matters when product managers review results too.

Starter is $0 a month with 1 GB of processed data, 10k scores and 14 days of retention; Pro is $249 a month with 5 GB, 50k scores and 30 days, then $3 per extra GB and $1.50 per 1,000 extra scores. On Starter, extra usage is billed too: $4 per GB and $2.50 per 1,000 scores.

In 2023, a Zapier engineer wrote in a developer forum that the company used Braintrust for evals.

Helicone is an AI gateway with LLM observability built in: teams use it to route, debug and analyze their AI apps. Setup is a base-URL change: point your OpenAI or other provider client at the Helicone gateway, add an API key, and every request is logged with cost, latency and tokens. The same gateway sits in front of every model provider you call.

Hobby is free with 10,000 requests and 7 days of retention; Pro is $79 a month with unlimited seats and a month of retention; Team is $799 a month with 3 months. Usage-based pricing applies on top. The code is Apache 2.0 licensed and self-hostable with Docker Compose.

The trade-off: request-level logging is its home ground. For deep agent traces and evaluation workflows, add Langfuse or Braintrust. Check the company's status too: a banner on its site says Helicone is no longer part of Mintlify, with details to come.

Confident AI is the cloud from the team behind DeepEval, the open-source evaluation library. It brings DeepEval's metrics to traces, datasets and regression tests, with a workflow built around catching quality drops before and after release.

A Free Forever plan covers 2 user seats, 1 project and 1 GB-month of trace spans, with 5 test runs a week. Paid plans are billed per organization: Starter is $200 a month with 5 GB-months of trace spans, Team is $2,000 a month with 75 GB-months, then $1 per GB-month. That flat team price works for a company with many users, less so for a solo developer.

Watch the free plan's ceiling: trace spans beyond 1 GB-month are dropped, data is kept for a week, and test runs beyond 5 a week are locked. Confident AI and Braintrust both publish comparison articles that rank on Google for this query, so read vendor rankings with that in mind.

Datadog's agent observability sits inside the Datadog platform, which is its whole argument: an LLM span links to the API request, the database query and the log line around it, and alerts use the monitors your on-call team already knows. Evaluations run on the same traces, and the Free tier already includes unlimited evaluations.

The Free tier includes up to 40K LLM spans a month with 15-day retention. Pro starts at $160 a month on an annual commitment for the first 100K spans, $200 month-to-month, or $240 on demand. Over a year on the annual rate, that is $1,920 before extra spans, which cost $3.50 per 10K on annual billing.

If you're not on Datadog already, the bill and the setup are hard to justify for LLM traces alone. Thinking of leaving it? See our Datadog alternatives.