Best LLM Observability Tools in 2026: Langfuse vs LangSmith vs Braintrust vs Arize Phoenix vs Helicone
The category split into three different jobs, and each tool bills for a different thing. Which one is cheapest depends entirely on the shape of your team and your traffic.
The agent worked in the demo. Then it shipped, someone asked it something slightly sideways, and it called the wrong tool three times before inventing a refund policy. Now you are looking at logs that show the final answer and nothing about how it got there.
That is the afternoon people go shopping for an observability tool. The problem is that the category quietly split into three different jobs, and most comparisons are written by one of the vendors and pretend their tool does all three equally well.
Prices below came off the vendor pricing pages on 21 August 2026.
Three jobs, and why nobody aces all three
Tracing is recording every step of a run: prompts, tool calls, retrieved chunks, latency at each hop, so you can replay a failure instead of guessing at it. This is what everyone needs first.
Evals are scoring outputs systematically, offline against a dataset before deploy or online against live traffic. This is how you catch “did my prompt edit make the summarizer worse.” Hardest job to do well, and the one most teams underuse.
Cost and latency is knowing which feature or customer or agent run is burning the tokens. Sounds trivial right up until the bill triples and nobody can say why.
No tool aces all three because they pull the architecture in different directions.
Langfuse: the open-source default
If a team does not yet know what it needs, this is what I would put in front of them. MIT licensed, genuinely full stack across tracing, evals, prompt management and datasets, and self-hostable with no feature gating. That combination is rarer than it sounds.
Framework-agnostic matters more than the phrase suggests. It does not care whether you are on LangChain, LlamaIndex, a raw provider SDK, or something homegrown. A decorator or an OpenTelemetry integration and you are done.
Cloud pricing: Hobby is free with 50,000 units a month, 30-day retention, 2 users. Core is $29 with 100,000 units and 90-day retention. Pro is $199 with three-year retention. Enterprise is $2,499. Overage is $8 per 100,000 units and every paid tier has unlimited users, which is a different philosophy from billing the people.
ClickHouse acquired Langfuse on 16 January 2026. The commitments made at the time were that it stays MIT, stays self-hostable, and the cloud keeps running as-is with no planned licensing change. So far that has held. It also tells you where the product is aimed: storing and querying very large volumes of trace data is exactly the problem ClickHouse exists for.
The catch is the one people always underestimate. Self-hosted Langfuse at production volume wants Postgres, a ClickHouse cluster, object storage, Redis, and app servers, plus somebody carrying the pager. Run that honestly and the total lands north of the cloud Pro tier. Self-host for data sovereignty, not to save money, because usually you will not.
LangSmith: correct if you live in LangChain
Made by the LangChain team, which is both the pitch and the caveat. On LangChain or LangGraph the tracing lights up with nearly no work, and the newer agent and eval features are designed hand in glove with the framework.
Pricing is per seat, and that is the thing to watch. Developer is free for one seat with up to 5,000 base traces a month. Plus is $39 per seat with 10,000 base traces, then pay as you go. Base traces are kept 14 days; extended retention is 400 days and costs extra.
Ten people on Plus is $390 a month in seats before a single trace overage. Langfuse Pro is $199 flat with unlimited users. That gap is not a discount war, it is two different opinions about what to meter.
My read: the right call when you are already committed to LangChain and value the integration over portability. If you are deliberately avoiding framework lock-in, note that you are not only picking an observability tool, you are deepening a bet.
Braintrust: when evals are the actual problem
Comes at this eval-first and it shows. If your question is not “help me debug this one trace” but “prove version B of this prompt beats version A across 500 cases,” this is the tool built for that. Datasets, scorers, experiments and side-by-side comparison are the product, not a tab bolted onto a tracing tool.
Starter is free with 1 GB of processed data, 10,000 scores a month and 14-day retention. Pro is $249 flat with unlimited users, 5 GB, 50,000 scores and 30-day retention. Overage on Pro is $3 per GB and $1.50 per 1,000 scores.
The billing detail that catches people: scores are metered separately from data and everything counts. An LLM-as-judge run, a heuristic check, a human clicking approve. Each one burns quota. Since the entire point of the tool is that you will run a lot of them, model score volume rather than traffic, because scores scale with your ambition.
It is the priciest entry point here and that is a useful filter. If you cannot yet say what would go in an eval dataset, you are not ready and Langfuse’s evals will cover you. If you already have a growing test set and eval is a daily ritual, $249 buys a workflow the generalists do not match.
Arize Phoenix: research-grade metrics, notebook-first
Phoenix comes out of the older ML observability world and feels like it, in a good way. It ships a deep library of research-backed evaluation metrics, which is what you want for serious retrieval relevance work or measuring hallucination with something more rigorous than asking a model whether the answer looks fine. Notebook-first, which suits people who already live in Jupyter.
Free to self-host, with one asterisk worth reading. It is under the Elastic License 2.0, not an OSI-approved license. ELv2 lets you use, modify and run it internally for commercial purposes but forbids offering Phoenix itself as a hosted service to third parties. For a team instrumenting their own app that never bites. For a platform company planning to resell observability it is a different deal from Langfuse’s MIT or Helicone’s open-source terms.
It connects upward into Arize’s enterprise platform when you outgrow the free piece. Reach for it when metric rigor matters more than dashboard polish, and when the people using it are comfortable without a point-and-click UI.
Helicone: cost visibility this afternoon
Helicone made the opposite architectural bet and for one specific need it is the smartest thing on the list. It is a proxy. Change your base URL, and every request flows through and gets logged with cost, latency and tokens. One line. No SDK, no decorators, nothing to instrument.
That bet is the whole story. The upside is the fastest possible path to cost and latency visibility across several providers, plus caching, retries and rate-limit handling for free because it already sits in the request path. The downside is that a proxy sees requests and responses, not the internal shape of a multi-step agent. You will not get “the planner called tool A, which returned X, which the model then reasoned about.”
Hobby is free with 10,000 requests a month on one seat. Pro is $79 with unlimited seats, alerts and reports, and usage-based billing above the included requests. Team is $799 and adds compliance and multiple organizations.
Use it when you need the dashboard today and do not yet need deep tracing or evals. It is also the right answer when you cannot modify the inference code at all: route traffic through a self-hosted instance and you get centralized logging without touching the server. Plenty of teams run Helicone for cost next to a second tool for tracing, and that is a reasonable stack rather than an admission of failure.
The number that actually decides your bill
Sticker price is the least interesting figure here. What decides the invoice is the unit each vendor chose to meter.
So do not compare $199 to $249 to $39. Estimate your traces per month, your headcount, and your eval score volume a year out, then run each vendor’s pricing against those three numbers. The winner is rarely the lowest headline.
Picking
Do not know yet, or want portability and the self-host option: Langfuse.
All in on LangChain or LangGraph: LangSmith, and watch the seat line as the team grows.
Evals are the central problem: Braintrust.
Research-grade metrics and a team that lives in notebooks: Phoenix, with the license read.
Cost and latency visibility before dinner: Helicone.
If you are starting from nothing, put self-hosted Langfuse or Helicone’s free tier in front of one real agent this week. Seeing your own tool calls and your own token counts settles which of the three jobs you actually have faster than any comparison table, this one included.