LangSmith alternatives

The shortlist

Prices and limits checked against each vendor's own site on 2 August 2026, and every one of those sites is linked at the foot of this page. Self-hosting is free of license cost, not free: you run and operate the stack, and that is a real line in the budget.

When to stay on LangSmith

If your agent is built on LangChain or LangGraph, the marginal cost of keeping LangSmith is close to one environment variable, and nothing in this roundup will be cheaper to keep than the thing that already works. Leaving over a feature you have not tried to configure is the most common way this goes badly.

Stay too if what you need is production monitoring at real volume, tracing in Go or Java, or SSO with SCIM and role controls. LangSmith Enterprise is the only tier here that runs self-hosted or hybrid inside your own VPC without giving up the hosted feature set, and a team that replaces it to save $39 a seat a month usually ends up rebuilding half of it.

And if you have neither observability nor a test environment today, start with observability. You cannot rebuild, replay, or score what you never recorded.

You want to self-host, or you want open source

This is the reason with the cleanest answer, and it is why most of these searches start. LangSmith will self-host, but only on Enterprise, so a small team that needs the traces to stay inside its own network has no priced path to it.

One detail worth being precise about, because it gets repeated wrongly: the Langfuse self-hosted build includes SSO and project-level role controls at no license cost, and its evals, annotation queues, and playground have not been behind a paid plan since June 2025. Self-hosting it does not get you a crippled edition.

You want a smaller bill

Two shapes of answer. Either you self-host and pay in operator time instead of subscription, which is everything in the section above, or you move to a hosted tier that meters something cheaper. On the hosted side the cheapest credible option is Langfuse Cloud, listed above: Hobby is $0 for 50,000 units a month across 2 users, and Core is $29 a month.

Before switching on price, work out what you are actually paying for. LangSmith meters base traces in normalized LangSmith Units: 1 LSU is $1.00 and a base trace is 0.005 LSU since the July 2026 repricing, with 14 days of base retention and 400 days extended. Whether that is expensive depends entirely on your trace volume and shape, so convert last month's real usage before assuming. A migration you did not need is the most expensive item on this page.

You want more metrics, or you want red teaming

LangSmith's eval surface is complete rather than deep: offline and online evals, LLM-as-judge, datasets, annotation queues, and CI through pytest and Vitest. Teams that leave over evaluation are usually not missing the framework, they are missing a library of metrics they did not have to write, or they have been handed a security question they cannot answer with any of it.

There is one answer to both, and it is the same product.

You want a sharper eval loop

This one is a taste judgement rather than a capability gap, and it is still a real reason to move. If the thing you do forty times a day is change a scorer and re-run, the friction of that loop is most of your working life.

The trade is what you pay for. Braintrust meters gigabytes of data processed and number of scores rather than traces or seats, which is not more expensive so much as differently shaped, and it is harder to forecast until you have run a month of your own volume through it.

Your problem is not scoring at all

There is a version of this search that none of the six options above answers, and it is worth naming because it costs people weeks. If you replace LangSmith and the thing you still cannot do is run the agent at all, because it calls an internal API nobody will give you a test instance of, or because every test run charges a real card or deletes a real record, then you do not have a scoring problem. Every product above scores a run that already happened somewhere, and somewhere is the part you are missing.

That is the gap Mirrors fills, and it is the only reason we are on our own roundup. Most readers of this page should skip this section: if your agent mostly reads, a RAG assistant, a summarizer, a classifier, then nothing it touches needs standing in for and the problem does not exist for you.

How to test an agent that calls internal APIs with no staging environment →

What moving actually costs

The migration is mostly instrumentation and history, in that order. Langfuse, Arize, and MLflow all accept OpenTelemetry, so if your tracing already goes through OTel the code change is an endpoint and a header. If it goes through the LangSmith SDK directly, budget a day per service to swap it, and run both for a week rather than cutting over.

History is the part people forget. Traces on the old platform generally do not move, and retention windows differ (LangSmith is 14 days at base, Braintrust 14 days on Starter, Arize AX 15 days on Free), so export anything you need for a baseline before you stop paying. Prompts, datasets, and annotations are usually exportable through each platform's API and are worth taking with you.

And whatever you land on, keep the eval set. The scorers are portable in principle and the cases are the asset: they encode what your team decided "good" means, and rewriting them is what actually takes the quarter.