LLM Observability Pricing: Braintrust vs Phoenix vs Langfuse

Verified as of 2026-06-16 against the official vendor pricing and documentation pages cited below.

The trap in this category is comparing a dollar figure across tools that meter different things. Braintrust charges for processed data in GB and for evaluation scores in thousands. Langfuse charges per unit, and a single request is rarely a single unit. Phoenix charges nothing to self-host and routes paid users to Arize AX at a flat $50 per month. Below, each pricing model is priced against the others on its own terms, with a head-to-head table at the end. The 2026 backdrop matters too: observability tooling is consolidating around OpenTelemetry tracing, which makes the billing unit, not the feature list, the real differentiator.

1. Three tools, three billing units

Braintrust bills on two axes at once: processed data measured in GB, and evaluation scores measured per 1,000. Its Starter tier includes 1 GB and 10,000 scores; overage is $4 per GB and $2.50 per 1,000 scores^[1]. A workload that runs many evals against the same traces hits the score ceiling first, while a high-volume logging workload hits the GB ceiling first.

Langfuse bills on one axis with a hidden multiplier. A unit is one trace, one observation, or one score, and the official formula is units equal traces plus observations plus scores^[3]. One user request is one trace plus however many spans and scores it generates, so the unit count per request is set by how deeply you instrument, not by traffic alone^[2].

Phoenix bills nothing for the open self-hosted build. It is licensed under the Elastic License 2.0 and runs locally or self-hosted with OpenTelemetry-based tracing, evals, and prompt management^[5]. The paid path is Arize AX, which meters spans and ingested GB rather than Langfuse-style units^[4].

2. Free tiers and where each breaks

Braintrust Starter is free with $10 of monthly credits, 1 GB of processed data, 10,000 scores, and 14-day retention; it allows one human-review score per project^[1]. It breaks the moment your eval volume crosses 10,000 scores or your data crosses 1 GB, at which point overage rates apply rather than a forced plan jump.

Langfuse Hobby is free with 50,000 units per month, 30-day retention, and two user seats^[2]. At six units per request, 50,000 units is roughly 8,000 instrumented requests a month; at twenty units per request for a deep agent, it is closer to 2,500. That spread is exactly why the free tier feels generous for simple apps and tight for agents.

Arize AX Free is genuinely free with 25,000 spans per month, 1 GB of ingestion, and 15-day retention^[4]. Phoenix self-hosted has no usage ceiling at all, only the compute you run it on^[5].

3. Braintrust: per GB plus per score

Braintrust positions itself around evaluation rather than passive logging, and its pricing reflects that. Pro is $249 per month and includes $249 of credits, 5 GB of processed data, 50,000 scores, 30-day retention, custom charts, and role-based access control^[1]. Overage on Pro is $3 per GB and $1.50 per 1,000 scores, both cheaper per unit than the Starter rates, so volume earns a discount.

The score axis is the one to watch. If you run an LLM-as-judge eval suite that scores every production trace on five dimensions, each trace consumes five scores, so 10,000 traces becomes 50,000 scores and exhausts the Pro allowance on scores alone before data volume matters. Beyond the included credits, Braintrust also meters its Topics feature at $0.06 per million input tokens and $0.40 per million output tokens^[1]. Plan around your eval depth, not your request count.

4. Langfuse: the per-unit multiplier trap

Langfuse Core is $29 per month with 100,000 included units, 90-day retention, and unlimited seats; Pro is $199 per month with 100,000 units and three-year retention; Enterprise is $2,499 per month^[2]. Overage is graduated: $8 per 100,000 units from 100k to 1M, $7 from 1M to 10M, $6.50 from 10M to 50M, and $6 above 50M^[2].

The estimation error people make is sizing a plan by request count. One request is one trace, but a request that runs three LLM calls and two evaluation scores is six units^[3]. A complex agent with retrieval, reranking, tool calls, and multi-step reasoning can run ten to thirty observations per request. At a midpoint of fifteen, 100,000 units covers roughly 6,600 requests, not 100,000. Langfuse-created data such as LLM-as-judge scores and experiment runs also counts as billable units, so your own evals inflate the meter^[3]. Estimate observations per request first, then multiply.

5. Phoenix: free to self-host, $50/mo for AX

Phoenix is the free-software option in this comparison. Under the Elastic License 2.0 you can self-host it via pip install and get OpenTelemetry tracing, response and retrieval evals, datasets, experiments, prompt management, and a playground at zero license cost^[5]. The one restriction that matters: the Elastic License forbids offering Phoenix as a hosted or managed service to others, which only affects you if you plan to resell it.

The managed counterpart is Arize AX. AX Pro is $50 per month with 50,000 spans, 10 GB of monthly ingestion, 30-day retention, and email support; extra spans bill at $3 per GB^[4]. So the real choice inside the Arize world is operational: run Phoenix yourself for free and own the infrastructure, or pay $50 per month for AX Pro and skip the hosting. For a solo founder who already runs a container or two, self-hosted Phoenix is the lowest-cost serious option in the category.

6. The verified comparison table

Dimension	Braintrust	Langfuse	Phoenix / Arize AX
Billing unit	GB processed + scores[1]	Units (trace + obs + score)[3]	Spans + GB ingested[4]
Free tier	$10 credits, 1 GB, 10k scores[1]	50k units/mo, 2 seats[2]	AX Free 25k spans; Phoenix self-host unlimited[4][5]
Entry paid plan	Pro $249/mo[1]	Core $29/mo[2]	AX Pro $50/mo[4]
Paid plan allowance	5 GB, 50k scores[1]	100k units[2]	50k spans, 10 GB[4]
Overage rate	$3/GB, $1.50 per 1k scores (Pro)[1]	$8 per 100k units (first tier)[2]	$3 per GB (AX Pro)[4]
Retention (entry paid)	30 days[1]	90 days[2]	30 days (AX Pro)[4]
Self-host free	No	Yes, open source[2]	Yes, Elastic License 2.0[5]

7. Which to pick

1. Pick Phoenix (self-hosted) if you want zero recurring software cost and can run a container. You get OpenTelemetry tracing and LLM-as-judge evals under the Elastic License 2.0, with no usage ceiling beyond your own compute^[5].
2. Pick Braintrust if evaluation is your center of gravity. Its score-metered model fits teams running structured eval suites, and Pro at $249/mo includes 50,000 scores; just plan for the score axis, not the GB axis^[1].
3. Pick Langfuse if you want a polished managed trace UI cheaply and your apps are not deeply nested. Core at $29/mo is the lowest managed entry price here, but estimate observations per request before you trust the unit allowance^[2].
4. Pick Arize AX if you want managed Phoenix without self-hosting. AX Pro at $50/mo with 50k spans is the middle path between free self-host and Braintrust's $249 floor^[4].
5. Either way, re-verify rates before committing; observability pricing shifts with each plan revision, and the billing unit, not the sticker price, decides your real bill.

For the API-layer cost that feeds these traces, the Anthropic vs OpenAI pricing and cheapest LLM API ranking articles price the model calls themselves. For the established tracing tools alongside these evaluation-focused ones, see LangSmith vs Langfuse vs Helicone.

Frequently asked questions

What does Braintrust cost per month in 2026?

Braintrust Starter is free with $10 monthly credits, 1 GB of processed data, 10,000 scores, and 14-day retention; overage runs $4 per GB and $2.50 per 1,000 scores. Pro is $249 per month with $249 of included credits, 5 GB processed data, 50,000 scores, 30-day retention, then $3 per GB and $1.50 per 1,000 scores over the limit. The two metered axes are processed data volume and evaluation score count, so a heavily-evaluated workload can hit the score ceiling well before the GB ceiling.

How are Langfuse units calculated?

A Langfuse billable unit is one trace, one observation, or one score. The official formula is units equal traces plus observations plus scores. A single request that runs three LLM calls and writes two evaluation scores is one trace, three observations, and two scores, which is six units. A complex agent with retrieval, reranking, and tool calls can reach ten to thirty observations per request, so request count alone undercounts your bill. Paid overage is $8 per 100,000 units on the first tier.

Is Arize Phoenix free?

Phoenix is free to self-host. It is licensed under the Elastic License 2.0, runs as a local or self-hosted web app via pip install, and gives you OpenTelemetry-based tracing, LLM-as-judge evals, datasets, and prompt management at no software cost. The Elastic License does forbid offering Phoenix as a hosted or managed service to third parties. The managed path is Arize AX, whose Pro tier is $50 per month.

Which LLM observability tool is cheapest for a solo founder?

Self-hosted Phoenix is cheapest at zero software cost if you can run a container. Among managed free tiers, Langfuse Hobby includes 50,000 units per month and Arize AX Free includes 25,000 spans, both genuinely free. Braintrust Starter is free with $10 of credits but its strength is evaluation rather than raw tracing volume. The right pick depends on whether your priority is eval workflows, trace volume, or zero recurring cost.