Weights & Biases Weave vs Langfuse vs LangSmith: Real LLM Observability Cost per 1 Million Traces in 2026

The short answer: these three don't meter observability the same way, and one of them doesn't meter traces at all. LangSmith bills per trace, with two retention tiers: 1 million traces, after the Plus plan's 10,000-trace included allowance, costs about $2,475 at the 14-day base-retention rate, or $4,950 at the 400-day extended-retention rate (roughly $2,500/$5,000 before accounting for the included allowance), plus a mandatory $39-per-seat-per-month platform fee on top, not included in either figure. Langfuse Cloud also bills per unit (a trace, an observation, or a score each count as one), but at a far lower rate: 1 million billable units costs about $101/month on its Core tier. That $101 figure is not directly comparable to LangSmith's per-trace cost, though: Langfuse bills a trace, an observation, and a score as three separate units, so one actual logged trace with several internal steps and scores can consume multiple billable units — a Langfuse workload that logs 1 million actual traces, each generating several observations, bills considerably more than $101 because each actual trace can generate multiple billable units. Weights & Biases Weave doesn't bill by trace count at all; it bills by gigabytes of data ingested, which means its cost depends entirely on how verbose your traces are, not how many of them you log — a structurally different meter that makes a direct "cost per 1 million traces" comparison impossible without first assuming an average trace size.

What each platform bills, and the billing-unit mismatch

LangSmith (OFFICIAL, LangChain's LangSmith pricing). Developer: free, 5,000 base traces/month, 1 seat. Plus: $39/seat/month, 10,000 base traces included. Beyond the included allowance, base traces (14-day retention) bill at $2.50 per 1,000, and extended traces (400-day retention) at $5.00 per 1,000 — a distinction this site's companion prompt-management-cost article covers in more detail for the related "Extended Data Retention Upgrade" mechanic that applies when a trace receives evaluation or annotation feedback. A separately cited, likely outdated figure of "$0.50 per 1,000 traces" appears in at least one source and conflicts with the $2.50/$5.00 structure found elsewhere; this article uses the more detailed, more recently verified rate.

Langfuse (OFFICIAL, Langfuse pricing; open-source core under MIT license). Bills in units, where a trace, an observation (an individual step within a trace), and a score each count as one unit. Hobby (Cloud): free, 50,000 units/month. Core: $29/month, 100,000 units included, then $8 per additional 100,000 units (effectively $0.08 per 1,000). Pro: $199/month. Enterprise: $2,499/month. Critically, Langfuse's entire core platform is open source and can be self-hosted for free, with no per-event cap of any kind on the self-hosted edition — the only cost at any volume is the infrastructure (Postgres, ClickHouse, Redis, S3) and the operational time to run it, which this article does not attempt to price since it varies enormously by team and existing infrastructure.

Weights & Biases Weave (OFFICIAL for the mechanism; REPORTED for exact current rates with a minor conflict). Bills by data ingestion volume, not trace count: Free: 1 GB of Weave trace ingestion a month. Team plans: $50 per user per month (one source cites $60/month for a comparable tier, a minor unreconciled conflict), including 1.5 GB of ingestion before per-MB overage applies. This is the same underlying mechanism as the broader W&B platform's seat-based pricing, with Weave-specific usage billed as an add-on. Because the meter is gigabytes, not traces, a verbose agent logging large prompts and full intermediate reasoning steps can consume its entire monthly allowance in far fewer traces than a lean, terse application — one industry account describes a team budgeting "$60 for W&B Pro" and instead seeing "a four-figure ingestion line" once real production trace volume and verbosity were accounted for.

Cost at 1 million LangSmith traces, versus 1 million Langfuse billable units

These are two different quantities, not a single "cost per 1 million traces" comparable across both platforms, because Langfuse's unit is not the same thing as a LangSmith trace.

The $101 Langfuse figure is valid for 1 million billable units — not necessarily 1 million actual traces, since each trace can generate multiple observations and scores that each count separately. The exact Langfuse cost for logging 1 million actual traces depends on how many observations and scores accompany each one, information this article does not have for any specific workload; a trace with many internal steps and evaluation scores could easily multiply the unit count, and therefore the dollar cost, several times over relative to the simplest possible trace shape.

Converting Weave's GB-based billing to a trace-equivalent cost

Formula: total GB = traces × average bytes per trace. At an ILLUSTRATIVE 20 KB per trace (a moderate-complexity agent trace with prompt, response, and a few intermediate steps), 1 million traces would ingest about 20 GB — far beyond even the Pro tier's 1.5 GB included allowance, pushing well into per-MB overage territory. This article could not find a confirmed per-MB or per-GB overage rate for Weave beyond the included allowance in the sources reviewed, so the exact dollar cost at this volume is QUOTE-ONLY / UNKNOWN; what is clear is that a verbose logging pattern multiplies Weave's cost directly, independent of how many distinct traces are logged, in a way that doesn't apply to LangSmith's or Langfuse's per-trace/per-unit meters.

The self-hosting escape hatch

Langfuse is the only one of the three with a genuinely free, fully-featured self-hosted option with no event cap — a real structural advantage for high-volume teams with the operational capacity to run their own Postgres, ClickHouse, Redis, and object storage. One industry account describes the typical arc as "everyone starts with hosted observability tools and eventually ends up building at least part of the stack themselves," citing Langfuse's self-host option specifically as the destination once a metered per-trace vendor's bill compounds past a team's comfort level. LangSmith has "no free self-hosted story" — self-hosting it in a private VPC is an Enterprise-tier feature with custom pricing, not a self-serve option. Weave's self-hosting is similarly limited to an enterprise-tier offering rather than a free, open-source path.

Sensitivity

1. Trace verbosity, for Weave specifically. The single largest and most unusual lever in this article; two teams logging the same trace count can have wildly different Weave bills depending on how much data each trace carries.
2. Which LangSmith retention tier applies. A full 2x cost difference between base (14-day) and extended (400-day) retention at the same trace volume.
3. Observations and scores per trace, for Langfuse specifically. Directly multiplies the billable-unit count, and therefore the dollar cost, for a given number of actual traces logged; this article's $101 figure assumes 1 million billable units, not 1 million traces with an unknown number of observations and scores each.
4. Self-hosting capacity. For a team with the operational ability to run Langfuse's open-source stack itself, the entire per-unit cost comparison in this article becomes moot, replaced by infrastructure and ops cost instead.
5. Seat count on LangSmith and Weave. Both carry mandatory per-seat platform fees on top of usage, which Langfuse Cloud's unit-based pricing does not include as a separate line.

Budgeting traps

• Treating Langfuse's $101-per-1-million-units figure as the cost of 1 million actual traces. A unit is a trace, an observation, or a score; a workload with several observations and scores per trace bills considerably more than this headline figure.
• Comparing Weave's per-GB pricing to LangSmith's or Langfuse's per-trace pricing without converting through an actual average trace size. The three are not priced in comparable units without that conversion.
• Assuming a $50 or $60/month Weave Team seat covers real production trace volume. Verbose agent traces can consume the included ingestion allowance far faster than trace count alone would suggest.
• Defaulting to LangSmith's extended retention tier when base retention would suffice. A 2x cost difference for data most teams don't need to keep for 400 days.
• Underestimating the operational cost of self-hosting Langfuse. The platform itself is free, but Postgres, ClickHouse, Redis, and S3 all need to be run and maintained by someone.

What to ask before you buy

Measure your actual average trace size in kilobytes before evaluating Weave's pricing, since its entire cost structure depends on data volume rather than trace count. Estimate your typical observations-and-scores-per-trace ratio before budgeting Langfuse off the $101-per-1-million-units headline, since that figure assumes 1 million billable units, not 1 million actual traces. Confirm which LangSmith retention tier (base or extended) your actual compliance or debugging needs require before defaulting to the pricier extended tier for all traces.