Agent observability platforms compared: licence, price, self-host

Choosing an agent observability platform looks like a feature comparison and usually is not. The question that eliminates most of the shortlist is simpler: are your prompts and outputs allowed to leave your infrastructure?

Traces from an agent contain the prompt, the tool arguments and the model’s response. That is customer data with extra steps.

Which platforms can you actually self-host?

Two, without a sales conversation.

If the answer to the residency question is no, the table is two rows long and the rest of this article is optional reading.

Note the licence column carefully. Langfuse is MIT outside its ee directories, but Arize Phoenix ships under the Elastic License 2.0, which is source-available rather than OSI open source. For internal use the practical difference is nil. If you intend to resell a managed service built on it, the difference is the whole conversation.

Running either one yourself also moves the bill from a subscription to a server, and whether that trade pays off depends on how busy the box is. The same break-even arithmetic that governs self-hosting a model applies to self-hosting your telemetry.

What the two pricing shapes cost as you grow

Why does the pricing model matter more than the price?

Because agent trace volume does not grow with headcount. It grows with runs, and runs grow with usage. LangSmith charges per seat and then per trace; Langfuse, Braintrust and Helicone charge a flat tier and then per unit of data. Those two shapes cross over somewhere, and where they cross decides your bill.

A seat-plus-usage model looks cheap during a pilot, when three engineers generate a few thousand traces. The same model at production volume is billing per thousand traces against a workload where a single task can emit dozens of spans. A ten-step agent producing one span per model call and one per tool call is emitting twenty-plus spans for one unit of user-visible work, and multi-agent designs multiply that again, for the same reason they multiply token spend.

Flat or self-hosted pricing looks expensive at the start and stops surprising you later. That is the actual trade, and it is decided by which growth curve your workload is on rather than by which number is smaller today.

The sampling question nobody asks early

Every usage-priced platform pushes you toward sampling, and sampling an agent trace is not like sampling a web request. Dropping 90% of HTTP requests still leaves a representative picture. Dropping 90% of agent runs means the specific failed run you need to investigate is probably gone.

Decide the retention and sampling policy before the pricing model decides it for you.

What does each one actually do well?

The categories are genuinely different, which is why the market has not consolidated.

• LangSmith has the deepest integration with LangChain and LangGraph, which matters if that is your framework. It is proprietary, self-hosting sits behind Enterprise, and its base traces are retained 14 days unless you pay to upgrade them to the 400-day tier.
• Langfuse is the open-source default: tracing and prompt management under MIT outside the ee directories, self-hostable, cheapest to start at $29 a month. Evaluation typically arrives through additional tools, so the stack tends to grow.
• Arize Phoenix is OTel-native with a strong evaluation metrics library, which suits teams that already treat OpenTelemetry as the standard rather than as one exporter among several. It carries 11.5k GitHub stars and the Elastic License 2.0.
• Braintrust is built around evaluation-driven development with gates in continuous integration, which is a different job from tracing and is why it is often used alongside rather than instead. Its Starter tier retains data 14 days; on-premise deployment is Enterprise-only.
• Datadog and Honeycomb are the pragmatic answer when your organisation already runs one of them. Agent spans in the same backend as everything else is worth more than any single feature.

How do you avoid paying for this decision twice?

Instrument to the OpenTelemetry semantic conventions for generative AI systems rather than to a vendor SDK.

Those conventions define the span names and attributes for agent, workflow, tool and model spans. Emit those and your traces are readable by any backend that accepts OpenTelemetry, which is all five platforms in the table above. Switching platform becomes a change of exporter endpoint rather than a project. The same span attributes also carry token counts, which is what lets you attribute cost to a specific step rather than to a monthly invoice.

Instrument to a vendor’s proprietary client and you have written the same telemetry in a dialect only that vendor reads. The cost of that decision is not visible until the day you want to leave, which is exactly when you have the least appetite for a re-instrumentation project.

The question to answer before you shortlist

Not “which platform has the best UI”. Ask: can prompts and outputs from production leave our infrastructure, and who signed off on that? The answer removes three names from the list in one step, and it is much cheaper to ask now than after the first trace has been sent.

Frequently asked questions

Method and sources

1. Langfuse pricing, langfuse.com/pricing: Hobby free at 50,000 units per month and 2 users, Core $29/mo, Pro $199/mo, Enterprise $2,499/mo. Checked 15 September 2026.
2. Langfuse repository, github.com/langfuse/langfuse: MIT licensed with the exception of the ee directories. Checked 15 September 2026.
3. LangSmith pricing, langchain.com/pricing-langsmith: Developer free with 5,000 base traces per month, Plus $39 per seat per month with 10,000 base traces, overage metered in LangChain Storage Units at $1.00 per LSU, base retention 14 days and extended retention 400 days. Checked 15 September 2026.
4. Arize Phoenix repository, github.com/Arize-ai/phoenix: Elastic License 2.0, OpenTelemetry-based instrumentation, 11.5k GitHub stars. Checked 15 September 2026.
5. Braintrust pricing, braintrust.dev/pricing: Starter free with 1 GB processed data and 10,000 scores per month at 14-day retention, Pro $249/mo, on-premise and hybrid deployment on Enterprise only. Checked 15 September 2026.
6. Helicone pricing, helicone.ai/pricing: Hobby free at 10,000 requests and 1 GB storage per month, Pro $79/mo, Team $799/mo, on-premise deployment on Enterprise only. Checked 15 September 2026.
7. OpenTelemetry GenAI semantic conventions, which cover spans, metrics and events for GenAI clients, MCP and provider-specific conventions, github.com/open-telemetry/semantic-conventions-genai. The conventions moved to this repository from the main semconv specification. Consulted for the portability argument in this article on 15 September 2026.
8. LangChain no longer publishes a flat per-1,000-trace overage rate; the $2.50 per 1,000 base traces used in the arithmetic below is the figure reported by 2026 third-party pricing write-ups, not a vendor-published rate. Treat the totals as illustration, not as a quote.
9. Vendor pricing changes frequently and should be re-checked before any purchase decision. We have not run all of these platforms side by side and do not present this as benchmark or performance data.