Agent Observability Compared: Langfuse, Helicone and Braintrust

Last reviewed: 9 October 2026. Pricing in this category changes often — verify against the vendor's page before you commit.

Method disclosure, up front — read this before the tables. We have not run all three of these tools side by side in production. This comparison is built from published pricing and documented capabilities as of 9 October 2026, plus the cost arithmetic those published figures imply at a realistic volume. Where sources disagreed, we say so instead of picking the convenient number. If you need a hands-on verdict on latency overhead or SDK ergonomics, this article will not give you one.

Affiliate disclosure: we hold no affiliate relationship with any of the three vendors at the time of writing and earn nothing from these links. If that changes, this line changes with it.

The short answer

If none of those four sentences describes you, the honest answer is that any of the three will do and the decision is not worth a week of evaluation. Pick the one whose free tier covers your current volume and revisit at 10×.

Langfuse

Shape: open-source, SDK-instrumented. Acquired by ClickHouse in January 2026; published capabilities unchanged since.

Pricing as published:

What it is good at: the self-hosted story is the differentiator and it is not a crippled edition — no per-seat pricing, intended for production. If your traces contain customer data, regulated data, or anything that makes a data-processing agreement a six-week project, this is the option that routes around the problem entirely.

What to check before committing: self-hosting is free in licence, not in operations. You are running the service, its database and its retention. Budget the engineer-hours; at small volume the cloud Pro plan is cheaper than the time you will spend on it.

Helicone

Shape: open-source, proxy-based. A one-line base-URL change routes your LLM calls through Helicone, which logs requests and responses and tracks cost, latency and usage.

Pricing as published:

What it is good at: time to first trace. The proxy model means you are not instrumenting call sites — you change a base URL and data appears. For a team that wants visibility today on an existing codebase, nothing else here is close.

What to check before committing: the proxy is in your request path. That is a latency consideration and an availability consideration — your inference now depends on their uptime unless you configure a bypass. Also note the retention tiers: 7 days on free is short enough that an incident you investigate on Monday may have aged out by the time you get to it. If compliance certification is a requirement, it lives on the $799 tier, not the $79 one.

Braintrust

Shape: eval-first platform with observability attached. No self-hosting.

Pricing — and here the published figures disagree. Two independent 2026 summaries report different tiers:

We are not going to resolve that by guessing. The discrepancy is probably a pricing change between the two publication dates, and the lesson is the practical one: verify the current tier on the vendor's own pricing page before you budget. A comparison article is a snapshot, including this one.

What it is good at: if your actual problem is "we do not know whether our agent got better after that change", Braintrust is built around that question rather than treating it as an add-on to logging.

What to check before committing: no self-hosting is a hard constraint, not a preference — if your data cannot leave your infrastructure, this tool is out regardless of how good it is. Also check the per-seat limits on the free tier; the two sources above differ on whether users are capped.

The arithmetic at volume

Free tiers mislead, because observability cost scales with traces and traces scale with success. At 500,000 traces per month with one score each, published figures work out to:

An 80 % premium, for a volume a moderately successful agent product reaches quickly. Before you dismiss 500k as hypothetical: a single agent run in a multi-turn workload emits dozens of spans. 500k traces is not 500k users — it can be a few thousand.

This is the number to compute for your projected volume, with your span-per-run ratio, before signing anything. It is also the number most teams discover after signing.

And the structural point: the self-hosted option is $0 in licence at any volume. The crossover where self-hosting beats cloud pricing is a function of your engineer-hour cost, and at 500k traces per month it has almost certainly already passed.

The dimension nobody puts in a comparison table

Three questions that decide this more often than feature lists do:

1. Proxy or SDK? Helicone sits in the request path; Langfuse and Braintrust instrument your code. The proxy is faster to adopt and harder to remove from your availability story. The SDK is more work up front and leaves your inference path untouched. This is an architectural choice, not a feature difference, and it is the one you cannot cheaply reverse.

2. Where does the data live? Agent traces contain prompts. Prompts contain whatever your users typed and whatever your retrieval layer pulled in. That is frequently the most sensitive data in the system, and it is routinely treated as debug output. Decide your data-residency answer before you choose a tool, because it eliminates options.

3. What will you actually look at? Most teams install observability, watch the dashboard for two weeks, then never open it again — because nothing is wired to a decision. Before choosing, name the three questions you want answered weekly (ours: cost per completed task, failure rate by tool, p95 turn count). If a tool cannot answer your three, its other four hundred features are decoration.

Our recommendation, stated plainly

• Default: Langfuse, self-hosted, if you have the operational capacity. It is the only option with no cost ceiling, no vendor in the request path, and no data-residency conversation.
• If you want traces this afternoon: Helicone, accepting the proxy in your path, and plan the retention tier deliberately.
• If evals are the actual problem: Braintrust, after confirming its current pricing and that no-self-hosting is acceptable.
• If you cannot decide: start on a free tier and instrument cost per completed task first. That single metric will tell you more in a month than a tool comparison will.

Sources

• Best LLM Observability Tools 2026: Langfuse, Helicone and LangSmith Compared — devtoollab
• Langfuse vs Helicone (2026): Which LLM Observability Tool to Pick
• Braintrust Alternative: Langfuse vs. Braintrust — Langfuse (vendor-published; treat the comparative framing accordingly)
• Braintrust Alternatives for LLM Eval & Observability — markaicode
• LLM Observability Tools: Langfuse vs Helicone vs Phoenix Compared
• Top 7 LLM Observability Tools in 2026 — Confident AI

Not tested hands-on. All pricing and capability claims are published figures as of 9 October 2026, cited above; one vendor's tiers are reported inconsistently across sources and we have flagged it rather than resolved it. The cost-at-volume figures are the sources' arithmetic on published rates, not our measurements.

← All articles