LangSmith Alternatives

Langfuse is the name that comes back most often when developers talk about leaving LangSmith, and it does the same three jobs: it records every step of an LLM call or agent run as a trace, lets you score those traces with evaluations, and keeps your prompts under version control. It plugs into LangChain through a callback handler that captures model calls, tools and retrievers, so a LangChain app can switch with a few lines of setup. It also takes traces from OpenTelemetry on its own endpoint, so code that is not built on LangChain works as well.

The pricing model is the main reason to move. LangSmith bills per seat and per trace; Langfuse bills per unit, where a unit is any tracing data point you send, such as a trace, an observation inside it or a score, and every paid plan has unlimited users. Hobby is free with 50k units a month, 30 days of data access and 2 users. Core costs $29 a month for 100k units and 90 days of data access, then $8 per 100k units up to 1 million, $7 per 100k from 1 to 10 million. Pro, at $199 a month, extends data access to 3 years and adds higher rate limits. The code outside the enterprise folders is under the MIT license, and ClickHouse, which bought Langfuse in 2026, says it plans no licensing change. If you already run it and want to compare further, see our Langfuse alternatives.

Arize Phoenix is the open, self-hosted product of Arize, the company behind the Arize AX cloud platform. Phoenix records traces of your LLM and agent runs, lets you run evaluations on them, and gives you datasets and experiments to compare prompt or model changes. It ships as Docker images and Helm charts, and Arize states that self-hosting comes with no license fees, no usage limits and no feature gates. That last point is the opposite of LangSmith, where self-hosting and hybrid deployment are reserved for the Enterprise plan at a custom price.

Read the license before you call it open source: Phoenix is under the Elastic License 2.0, which lets you use and modify it inside your company but forbids offering it to others as a managed service. For a team that runs it for itself, that changes nothing. If you would rather not host anything, Arize AX is the cloud route. AX Free gives 25k spans and 1 GB of ingestion a month; AX Pro costs $50 a month for 50k spans and 10 GB. A span is one step inside a trace, so with ten steps per run, AX Pro covers about 5,000 runs a month, half of what a single LangSmith Plus seat includes. Arize does not publish a price per extra span on that page, so ask before you pass the quota.

Braintrust starts from evaluation rather than from tracing. You build datasets, write scorers, run experiments on each prompt or model change and compare the results side by side; production logs feed back into the same datasets. It also has Loop, an assistant inside the product that helps you write scorers and improve prompts. Teams that use LangSmith mainly for its evaluation features, not for LangChain debugging, often find Braintrust closer to how they work.

The pricing unit is data, not traces or seats, and every plan includes unlimited users. Starter is free with 1 GB of processed data, 10k scores and 14 days of retention, then $4 per GB and $2.50 per 1,000 scores. Pro costs $249 a month for 5 GB, 50k scores and 30 days of retention, then $3 per GB, $1.50 per 1,000 scores and $0.50 per GB per month for longer retention. That flat fee is high for a small team: three LangSmith Plus seats cost $117 a month. It becomes cheap when ten or twenty people need access. The platform itself is closed source, its SDKs are under Apache 2.0, and only Enterprise adds a bring-your-own-cloud deployment that keeps the data in your own cloud; Starter and Pro run on Braintrust's SaaS. Below Enterprise, single sign-on is Google only. If Braintrust is already on your list, our Braintrust alternatives show where its users go.

Opik is Comet's open source platform for tracing, evaluating and monitoring LLM apps and agents. It covers what most teams use LangSmith for: traces of each run, LLM-as-a-judge evaluations with more than 30 ready metrics, datasets and experiments, prompt management, plus an agent optimizer that tunes prompts for you, and, on the self-hosted version, guardrails that block PII and unwanted topics. Its code is under the Apache 2.0 license, a permissive license like Langfuse's MIT core, so you can run it, change it and even host it for others.

The hosted plans are the cheapest here. Free Cloud covers up to 10 team members and 25k spans a month. Pro Cloud costs $19 a month for up to 50 team members, 100k spans and 60 days of retention, then $5 per 100k extra spans. Keeping spans longer than 60 days costs $29 per 100k spans, almost six times the price of ingesting them, so set retention before you raise it. A self-hosted Opik has unlimited spans and retention, and it is where the guardrails live: Comet offers them for self-hosted deployment, not on Free Cloud, Pro Cloud or Enterprise. It also comes without user management, which matters if several teams share one instance. The Free plan cannot buy extra spans.

Weave is the LLM tracing and evaluation product of Weights & Biases, now part of CoreWeave. You add a decorator to the functions you want to follow, and Weave records inputs, outputs, cost and latency for each call, then lets you build evaluations and compare versions of a prompt or model. Its strength is that it lives next to W&B experiment tracking: a team that fine-tunes models and then ships them in an app keeps training runs, model versions and production traces in one place, which LangSmith does not cover.

Pricing is by data volume. The Free plan costs $0 a month with 1 GB of Weave data ingestion and 5 GB of storage. Pro starts at $60 a month billed monthly, with 1.5 GB of Weave ingestion and 100 GB of storage, then extra ingestion billed per megabyte, $0.10 per MB on Pro, which works out to $100 per extra GB. That rate is the trap: a chatty agent that logs long prompts and documents can pass 1.5 GB quickly, so check your average trace size first. Volume discounts come with annual commitments. One more limit sits in the plan description: Pro is meant for early-stage teams of fewer than 50 employees, and a company above that is required to move to the Enterprise plan, priced on request. The Weave SDK is under Apache 2.0, but the W&B platform you send data to is a commercial service.

Datadog LLM Observability adds LLM and agent tracing to the monitoring platform many engineering teams already pay for. Each model call shows up as a span you can correlate with APM services, infrastructure signals and real user sessions, so a slow answer can be traced to a database or a network issue, not only to the prompt. It also runs out-of-the-box checks for hallucinations and prompt injection attempts, and lets you run evaluations and experiments. For a platform team, the gain is one tool, one on-call process and one bill instead of LangSmith next to Datadog.

The free tier covers up to 40K LLM spans a month with 15 days of retention and unlimited evaluations. Pro starts at $160 a month on annual billing, or $240 on-demand, with 100K LLM spans; extra spans cost $3.50 per 10K on annual billing, $4.20 month-to-month and $5 on-demand. Only spans that call a model are billed, not tool, retrieval or workflow spans, which favors agents with many tool steps. Retention is where the bill grows: 30 days costs $6.50 per 10K spans a month and 90 days $9, more than twice the span price itself. Our Datadog switch guide covers the wider platform and why some teams leave it.