Best Langfuse Alternatives in 2026: 12 LLM Observability Tools Compared

By Camellia · Updated Nov 3, 2026

If you want what Langfuse gives you (open source, self-hostable, OpenTelemetry-friendly) under a different owner, start with Comet Opik, Arize Phoenix, MLflow, or OpenLIT. If you want evaluation to be the main job, look at Braintrust or Splunk Agent Observability (the product formerly called Galileo). If your team already lives inside LangChain or Datadog, LangSmith and Datadog LLM Observability will save you the most setup. This guide covers 12 options, with prices taken from each vendor's own pricing page and a cost table at one million units a month so you aren't comparing free tiers.

Langfuse is still a strong product, and most people searching for alternatives aren't running away from it. The usual reasons are narrower. ClickHouse announced it was acquiring Langfuse in January 2026, and some buyers want a different owner or a different license. Others hit the jump from the $199 Pro plan to the $2,499 Enterprise plan. A few need a gateway, a proxy, or APM correlation that Langfuse doesn't try to be. If your honest answer is "none of the above," stay put. The Langfuse vs LangSmith comparison covers the closest head-to-head, and the broader LLM observability roundup ranks the whole category.

Key Facts

• 89% of organizations surveyed already have some form of observability for their agents, but only 52.4% run offline evaluations and 37.3% run online evaluations, per LangChain's State of Agent Engineering report (1,340 responses, surveyed November 18 to December 2, 2025).
• 57% of those respondents have agents in production, and another 30.4% are actively building toward it, per the same report.
• 28% of LLM call spans carry cached-read input tokens even though the models in question support prompt caching, per Datadog's State of AI Engineering report covering February and March 2026.
• Rate limits caused 60% of the errors in Datadog's February 2026 data, while 5% of LLM call spans errored that month, per the same report. Those are failures a trace shows and a transcript hides.
• ClickHouse raised $400 million on January 16, 2026 and announced the Langfuse acquisition in the same post. Langfuse had reported more than 20,000 GitHub stars and over 26 million SDK installs a month at the end of 2025, per that announcement.

Quick Comparison Table

Why Teams Look Past Langfuse

Before shortlisting, write down which of these reasons is actually yours. Different reasons point to different tools.

What Langfuse Gives You, So You Know What You're Replacing

Langfuse's own pricing page lists four plans: Hobby is free (50k units, 2 users, 30 days of data access), Core is $29 a month (100k units, unlimited users, 90 days), Pro is $199 a month (100k units, 3 years of data access) and Enterprise is $2,499 a month. Overage runs $8 per 100k units from 100k to 1M, $7 from 1M to 10M, $6.50 from 10M to 50M and $6 above that, per langfuse.com/pricing.

The repository is MIT licensed except for the ee folders, and self-hosting runs on Docker Compose, a single VM, or Kubernetes with Helm, per the Langfuse GitHub README. ClickHouse says Langfuse stays open source under its MIT license for core features, that production-scale self-hosting continues, and that Langfuse Cloud keeps running as a standalone service, per its announcement. So the practical question isn't whether Langfuse works. It's whether you'd rather buy from someone else.

Who Owns Each Vendor Now

Ownership changed for six of these twelve in 2026 alone. Here is the confirmed state from each company's own announcement.

Two of the confirming pages give no closing date, so this guide doesn't either. Dynatrace's own post, published October 1, 2026 and updated October 21, says the acquisition is complete, and Splunk's page says Cisco "has completed its acquisition of Galileo."

OpenTelemetry, Self-Hosting and License at a Glance

Instrumenting once in OpenTelemetry and swapping backends later is the cheapest insurance against another acquisition. Self-hosting matters because a trace usually carries the full prompt and completion.

The licensing detail that trips people up: Phoenix describes itself as licensed under the Elastic License 2.0, per its GitHub page. That's source-available, not the permissive license Langfuse's core uses. If your legal team needs MIT or Apache for the whole platform, Opik, MLflow and OpenLIT are the clean answers.

1. LangSmith: Deepest Tracing if You Already Use LangChain

LangSmith is LangChain's tracing, evaluation and deployment platform. Point it at a LangChain or LangGraph app and runs show up as a full execution tree without extra instrumentation. It also takes OpenTelemetry traces, so a mixed stack can still land in one place. It's the most direct commercial rival to Langfuse, and the LangChain alternatives guide covers the framework side of that same ecosystem.

The catch is billing. Overage is metered in LangChain Standard Units at $1.00 each, and the pricing page doesn't state a flat per-trace rate, so a one-million-trace month can't be computed from published numbers.

Pricing: Developer $0 per seat (5k base traces a month, 1 seat). Plus $39 per seat per month (10k base traces, unlimited seats, one free small serverless deployment). Developer and Plus are billed monthly; Enterprise is invoiced annually upfront. Source: langchain.com/pricing. LangSmith receives OpenTelemetry spans at an OTLP endpoint.

Best for: teams standardized on LangChain or LangGraph. Not ideal for: teams that want one predictable per-trace number.

2. Arize Phoenix and AX: Free Self-Hosted Start, New Parent

Arize has two halves. Phoenix is the open-source platform: OpenTelemetry-based tracing, LLM evals, versioned datasets and experiments, a prompt playground, and a remote MCP server so MCP clients can query traces and datasets. It runs locally, in Docker, on Helm, or through one-click deploys on Railway, Render, Google Cloud Run, Azure and AWS, per the Phoenix repository. AX is the hosted commercial product.

Dynatrace's post says Arize AX and the Dynatrace platform remain standalone offerings and Phoenix remains an open-source project, per Dynatrace. It states no pricing, so confirm AX terms before committing past a pilot.

Pricing: AX Free $0 (25k spans a month, 1 GB ingestion, 15-day retention, unlimited users). AX Pro $50 a month (50k spans, 10 GB, 30-day retention, unlimited users). AX Enterprise custom, SaaS or self-hosted. The page describes no overage charges. Source: arize.com/pricing.

Best for: teams that want a free OpenTelemetry-based tool with an enterprise path. Not ideal for: buyers who need a permissive license on the whole platform.

3. Braintrust: Evaluation as the Center of the Product

Braintrust treats evaluation as the main job. Users, projects, playgrounds, experiments and datasets are unlimited on every tier, so you can run eval suites in CI before a prompt or model change ships. Tracing is present but isn't the headline.

Billing doesn't map to traces at all. You pay for processed data and scores, which makes it hard to compare against a per-unit competitor without your own volume estimate.

Pricing: Starter $0 per month (1 GB processed data, 10K scores, 14-day retention; overage $4 per GB and $2.50 per 1,000 scores). Pro $249 per month (5 GB, 50K scores, 30-day retention; overage $3 per GB and $1.50 per 1,000 scores; retention beyond 30 days $0.50 per GB per month). Enterprise custom. Source: braintrust.dev/pricing.

Best for: teams gating releases on eval scores. Not ideal for: teams whose main need is a deep trace viewer. For a three-way view see Arize vs Galileo vs Braintrust.

4. Comet Opik: The Closest Open-Source Match

Opik is the most direct open-source substitute. The repository is Apache 2.0 and its README says the full platform is free to self-host, via Docker Compose for local work and Helm on Kubernetes for production, per the Opik repository. It covers tracing for LLM calls and multi-step agents, datasets and experiments, LLM-as-a-judge metrics such as hallucination detection, a PyTest integration for CI, dashboards, online evaluation rules and a prompt playground.

Hosted pricing is low and the overage is public, which makes Opik the cheapest option in the one-million-unit table below.

Pricing: Open Source free, self-hosted. Free Cloud $0 (up to 10 team members, 25k spans a month, 60-day retention). Pro Cloud $19 a month (up to 50 members, 100k spans, 60-day retention; $5 per additional 100k spans). Extending retention from 60 to 400 days is $29 per 100k spans. Enterprise custom. Source: comet.com/site/pricing.

Best for: teams that want Langfuse's shape with Apache 2.0 and a cheap hosted tier. Not ideal for: buyers who weight ecosystem size above everything.

5. MLflow: Observability Inside an ML Platform You May Already Run

MLflow's repository calls it an open-source AI engineering platform for agents, LLMs and ML models. It captures traces built on OpenTelemetry, lists one-line tracing for more than 60 frameworks, and ships 50+ built-in metrics and LLM judges for evaluation, per the MLflow repository. It's Apache 2.0 licensed.

It isn't a drop-in Langfuse replacement in feel. If you already use MLflow for experiment tracking and model registry, adding GenAI tracing is a small step. If you start from nothing, you're adopting a much bigger platform than you need.

Pricing: Free and open source. Costs are your own hosting, or a managed provider you choose separately. Source: github.com/mlflow/mlflow.

Best for: ML teams already on MLflow. Not ideal for: a two-person team wanting a light trace viewer.

6. OpenLIT: OpenTelemetry-Native With GPU Monitoring

OpenLIT describes itself as an open-source platform for tracing, evaluating and improving AI agents. It follows OpenTelemetry GenAI conventions, accepts OTLP over gRPC and HTTP, and can export to other OTLP backends. Beyond traces it offers GPU monitoring for NVIDIA, AMD and Intel hardware used for inference, runtime guardrails for prompt injection and jailbreak attempts, and a versioned Prompt Hub, per the OpenLIT repository.

It's Apache 2.0, self-hosting is free with no license key and no per-trace fee, and it runs on Docker Compose or Kubernetes. The project is much smaller than Langfuse (about 2.9k stars), so plan for community support.

Pricing: Free, self-hosted. Source: github.com/openlit/openlit.

Best for: teams running their own models who want GPU and trace data together. Not ideal for: buyers who need a vendor-backed SLA.

7. Helicone: Fast Proxy Setup, Uncertain Roadmap

Helicone's design is a proxy: route calls through it and logging starts with almost no code change. Mintlify announced it was acquiring Helicone on March 3, 2026, and said it would continue in maintenance mode with security updates, bug fixes and new model support, while helping customers migrate to other platforms, per Mintlify. That is a signal to treat it as a quick proxy, not a platform to build on for years.

Pricing: Hobby free (10,000 requests, 1 seat, 7-day retention). Pro $79 a month (unlimited seats, 1-month retention, usage-based pricing beyond 10K free requests). Team $799 a month (unlimited seats, 3-month retention). Enterprise by contact. No per-request overage rate is published. Source: helicone.ai/pricing.

Best for: a fast proxy for cost and latency visibility. Not ideal for: anyone who needs an actively developed roadmap.

8. W&B Weave: Natural Fit for Weights and Biases Teams

Weave is the LLM and agent tracing product from Weights and Biases, and its pricing now sits on CoreWeave's Forge page. If your ML team already tracks training runs in W&B, putting traces and evals in the same place avoids another login.

Billing is by ingested data, not trace count: Pro includes 1.5 GB of Weave ingestion a month and charges $0.10 per MB beyond that, so payload size drives the bill. The page also says Pro is for organizations with fewer than 50 employees.

Pricing: Free $0 (up to 5 seats, 1 GB Weave ingestion, 5 GB storage a month). Pro starts at $60 a month billed monthly (up to 10 seats, 1.5 GB ingestion, 100 GB storage; $0.10 per MB ingestion overage, $0.03 per GB storage overage). Enterprise custom, invoiced annually, ingestion overage $0.20 per MB. Source: coreweave.com/forge-pricing.

Best for: ML teams already in W&B. Not ideal for: teams that want a per-trace price.

9. Datadog LLM Observability: Traces Next to Your APM

Datadog's pitch is correlation. If infrastructure, APM and logs already live there, LLM traces sit beside the same services. Only LLM spans (model calls) are billed; tool, workflow, agent, embedding and retrieval spans are free.

Pricing: Free $0 (up to 40K LLM spans, 15-day retention). Pro $160 a month on annual billing, $200 month-to-month, $240 on-demand, each including 100K LLM spans. Additional spans cost $3.50 per 10K annual, $4.20 month-to-month, $5.00 on-demand. Source: datadoghq.com/pricing.

Best for: platform teams already paying for Datadog. Not ideal for: teams that need self-hosting.

10. Splunk Agent Observability (Formerly Galileo): Judge Models and Guardrails

Splunk's page says Splunk Agent Observability is the new name for Galileo, which Cisco acquired, and describes helping teams evaluate agent behavior, observe performance across the stack and apply guardrail actions in real time across all traffic, per Splunk.

Buying has changed from a self-serve tier to annual contracts. The pricing page lists a free edition without limits, and paid pricing from $100 a month covering 1.2 million spans, billed annually. Each year includes 500 million Luna tokens and 5 Luna fine-tuning runs. Anything past that is by quote.

Pricing: Free edition available. Paid from $100 per month for 1.2 million spans, billed annually. Source: splunk.com pricing.

Best for: enterprises already on Splunk or Cisco. Not ideal for: monthly self-serve buyers.

11. Portkey: Gateway and Logs in One Bill

Portkey's core is an AI gateway for routing, fallbacks, load balancing and retries, with observability on the traffic that flows through. Palo Alto Networks completed its acquisition on May 29, 2026 and Portkey is now the core AI Gateway for Prisma AIRS, per the press release. The open-source gateway is free to self-host. For gateway-specific choices see the MCP gateways roundup.

Pricing: Open Source free, self-hosted. Developer free forever (10k recorded logs a month, 3-day log retention, 30-day metrics). Production $49 a month (100k recorded logs, 30-day logs, 90-day metrics; overage listed as $9 per additional 100k requests). Enterprise custom. Source: portkey.ai/pricing.

Best for: teams that need routing and logs together. Not ideal for: teams that need deep scoring.

12. Traceloop (OpenLLMetry): Own the Instrumentation Layer

OpenLLMetry is the open-source instrumentation library. It works with any OpenTelemetry backend, including several tools above. Traceloop announced in March 2026 that it is joining ServiceNow, that its technology becomes part of ServiceNow's AI Control Tower, that OpenLLMetry will remain open source, and that roadmap changes will be communicated in advance, per Traceloop.

Pricing: Free Forever $0 (up to 50K spans a month, up to 5 seats, 24-hour retention). Enterprise custom, with SOC 2 and on-prem deployment. The page's FAQ and its pricing card disagree on the seat limit. Source: traceloop.com/pricing.

Best for: teams that want instrumentation they can point anywhere. Not ideal for: buyers who want a self-serve paid plan.

What 1 Million Units a Month Costs

Free tiers don't tell you much. The unit differs by vendor, so read that column first.

The pattern: Opik and Langfuse stay cheap because their overage math is public and linear. Five vendors publish no number you can extend. That alone tells you who they're built for.

Stage and Sizing Fit

Memory, RAG and Agents: Adjacent Tooling

Observability is one layer. If the trace you're chasing is a retrieval failure, the RAG tools roundup covers the stack those traces record, and the agent memory tools roundup covers state that spans sessions. If you're tracing multi-step agent loops rather than single calls, the AI agent observability roundup frames the buying questions differently.

How to Choose: Decision Framework

Frequently Asked Questions

What to Do Next

Pick two candidates from the decision table and send the same real workload to both for two weeks. Check that you can see the full prompt and completion on a failed run, run one offline evaluation, and price the result at your actual monthly volume using the table above. If one tool's price at your volume isn't published, that's your answer about whether it fits a self-serve budget. For the wider category, the LLM observability roundup ranks all of these side by side.