Langfuse Alternatives

Arize Phoenix is the closest thing to Langfuse that is easier to run. It is built on OpenTelemetry and the OpenInference conventions, so the same instrumentation sends traces from LangChain, LlamaIndex, the OpenAI SDK, DSPy or your own code. In the UI you get the trace tree of each request, span-level evaluations with LLM judges, datasets, experiments that compare two prompt versions on the same examples, and a prompt playground. Arize states that self-hosted Phoenix has no license fees, no usage limits and no feature gates, and it ships as a Docker image, a Helm chart or a simple install from the terminal. That is the main reason teams pick it over Langfuse: one container plus a SQL database (SQLite or PostgreSQL) instead of six moving parts.

The catch is the license. Phoenix is under the Elastic License 2.0, which lets you use and self-host it freely but forbids offering it to others as a managed service. When you outgrow your own server, Arize sells a managed product called AX: AX Free gives 25,000 spans and 1 GB of ingestion a month, and AX Pro costs $50 a month for 50,000 spans and 10 GB. AX Free keeps data 15 days and AX Pro 30 days. AX counts spans rather than every data point, so a 30-step agent run uses about the same quota in both tools, but AX Pro includes half of Langfuse Core's 100,000 units for $21 more: about 1,666 such runs for $50 against about 3,225 for $29. Arize does not publish a price per extra span, and SSO, audit logs and self-hosted AX are Enterprise only.

LangSmith is the observability and evaluation platform made by the LangChain team. If your code already uses LangChain or LangGraph, you set a few environment variables (tracing on, an API key and, in the newer agent workspaces, the agent and its environment) and every chain, tool call and agent step appears as a nested trace, with no extra decorators. It also accepts traces from other frameworks and plain SDK calls, so it is not locked to LangChain, but that is where it saves the most time; our LangChain review covers the framework itself, and our LangChain alternatives page helps if you are not sure you will keep it. Around tracing you get datasets, offline and online evaluations, annotation queues for human review, a prompt hub and monitoring dashboards. The same account also sells agent deployment and other agent products, metered in LangChain Standard Units (LSU) at $1 each.

Pricing is per seat plus traces. Developer is free for 1 seat with 5,000 base traces a month. Plus costs $39 per seat a month, billed monthly, and includes 10,000 traces in total, then 0.005 LSU, or half a cent, for each additional trace. Base traces are kept 14 days; extended traces are kept 180 days for an extra fee. Three engineers sending 50,000 traces a month pay about $317: $117 in seats and $200 in extra traces. LangSmith gets expensive with headcount and volume. Two traps: the free Developer plan is pay-as-you-go above 5,000 traces rather than capped, and any automation rule that matches one run upgrades the whole trace to extended retention, at the higher price. Self-hosting is an add-on to the Enterprise plan.

Opik is the LLM observability and evaluation tool from Comet, the company known for ML experiment tracking. It covers the same ground as Langfuse: traces and spans, cost and latency, prompt versions, datasets, LLM-as-a-judge metrics, online evaluation rules and experiment comparison. Integrations cover OpenAI, Anthropic, LangChain, LlamaIndex, CrewAI, DSPy and OpenTelemetry. The repository is under Apache 2.0 with no enterprise folder carved out, which makes it the cleanest license of the self-hosted options on this page, and Comet's open-source column lists unlimited spans and unlimited retention.

The cloud plans are cheap. Free Cloud gives 25,000 spans a month for up to 10 team members, with 60 days of retention. Pro Cloud costs $19 a month for up to 50 members and 100,000 spans, then $5 per 100,000 extra spans; longer retention costs $29 per 100,000 spans. One million spans a month on Pro comes to $64. The same million units on Langfuse Core costs $101, so a team that moves to Opik's cloud saves money at volume and gets more seats on the free tier. Watch three limits: Free Cloud cannot buy extra spans, Free and Pro Cloud keep data in the US only, and self-hosted Opik comes without user management. Self-hosting is also no lighter than Langfuse: it runs ClickHouse, MySQL and Redis, and the production install targets Kubernetes.

Braintrust starts from evaluation rather than logging. You define a dataset, a task and scorers, then each run becomes an experiment you can compare with the last one, line by line, in the UI or in your CI. It adds a playground for prompts across models, the Braintrust Gateway (formerly its AI proxy, free while in public preview) that calls many providers through one API (our AI gateway ranking compares dedicated gateways), and Loop, an assistant that writes scorers and datasets for you.

Pricing is by processed data and scores, not by trace. Starter is free with 1 GB of processed data, 10,000 scores and 14 days of retention, then $4 per GB and $2.50 per 1,000 scores. Pro costs $249 a month with 5 GB, 50,000 scores and 30 days of retention, then $3 per GB, $1.50 per 1,000 scores and $0.50 per GB per month to keep data longer. Against Langfuse Pro at $199, Braintrust costs more and keeps data far less long: 30 days by default, 180 days at most with paid retention, versus 3 years of access. SSO is Google only below Enterprise. The platform is closed source, though its SDKs are open source; to keep data in your own cloud account, the Enterprise plan offers bring-your-own-cloud, with Braintrust operating the deployment. New customers at Series A or earlier can apply for 6 to 12 months of Pro for free.

Logfire comes from the team behind Pydantic. It is a general observability tool built on OpenTelemetry, with strong support for LLM calls: Pydantic AI, OpenAI, Anthropic and other SDKs are instrumented in one line, and the agent run sits in the same trace as the HTTP request, the database query and the background job around it. That is the gap it fills next to Langfuse, which by default keeps only LLM-related spans. You query your data with SQL, build dashboards and alerts, and it now adds evals. The Python SDK is MIT licensed, but the server and UI are closed source; self-hosting is an enterprise option priced on request.

Among the cloud plans on this page, its pricing is the most generous for volume. Every plan includes 10 million records a month (logs, spans or metrics). Personal is free for 1 seat and 2 read-only guests, and is hard-capped: ingestion pauses once you hit the 10 million records. Team costs $49 a month with 5 seats included, up to 12, $25 per extra seat and $2 per million extra records up to a spending cap you request by email; past the cap, new data is hidden until you raise it. Growth costs $249 a month with unlimited seats. Fifty million records a month on Team comes to $129. The limit for a Langfuse user is history: retention is 30 days on Personal and Team and 90 days at most on Growth, and evals and human review start with Team.

Weave is the LLM part of Weights & Biases, now owned by CoreWeave and sold under its CoreWeave Forge plans. You decorate your functions with weave.op, or use one of its integrations, and each call is logged with inputs, outputs, cost and latency in a trace tree. Evaluations compare models or prompts on a dataset with scorers, and leaderboards show the results side by side. The SDK is under Apache 2.0. Its real advantage is context: if your team fine-tunes models or tracks experiments in W&B Models, the LLM traces live next to the training runs, with the same projects, permissions and reports.

The free plan includes 1 GB of Weave data ingestion a month 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, and volume discounts on annual commitments. Two catches: past the included 1.5 GB, Pro bills extra ingestion at $0.10 per MB, so $100 per GB, and Pro is meant for organizations under 50 employees. Weave bills by data size, not by trace or span, so a verbose agent that logs long prompts uses its quota faster than a short chatbot. Self-hosting is an enterprise option, and teams that do not use W&B for anything else gain little from the bundle.

Datadog LLM Observability, now marketed as Agent Observability, adds LLM traces to the monitoring platform many companies already use. Each agent run shows its model calls, tool calls and retrievals, linked to the APM trace of the service that made them, so an on-call engineer sees the slow database and the slow model in the same view. You get out-of-the-box checks for hallucination, prompt injection and sensitive data, custom evaluations, experiments and cost tracking, plus all of Datadog's alerting, dashboards and role controls.

Pricing is per LLM span: only the spans that call a model are billed, not tool, retrieval or workflow spans. The free tier covers up to 40,000 LLM spans a month with 15 days of retention. Pro starts at $160 a month on annual billing for 100,000 spans ($240 a month on demand), then $3.50 per 10,000 spans on annual billing, $4.20 month to month and $5 on demand. One million spans a month on annual billing comes to $475. Retention is 15 days by default; 30 days costs $6.50 per 10,000 spans a month, and 90 days $9. Against Langfuse Core, where one million units costs $101 and keeps 90 days, Datadog costs several times more per billed unit. Experiment results are kept 15 days by default, 3 months with an annual or month-to-month commitment. If the Datadog bill as a whole is the problem, our guide to switching from Datadog covers the wider move.