LangSmith vs Braintrust vs Arize AX: Real LLM Observability Cost per 1 Million Traces in 2026

The short answer: the three platforms do not meter the same thing, so "cost per trace" only means something once you know whether the vendor counts traces, spans, or gigabytes. LangSmith bills per trace regardless of how many steps are inside it, so a 2-span chat request and an 18-span agent run cost the same. Arize AX bills per span and per gigabyte ingested, so the same agent run can cost eight times more than the chat request. Braintrust bills per gigabyte of processed data and per score, so its price follows payload size, not trace count. At 1 million traces a month, that produces roughly $2,592 on LangSmith regardless of shape, $324 to $333 on Braintrust Pro (before scoring volume), and $1,610 to $14,410 on Arize AX depending on whether the traces are simple chat calls or 18-span agent runs.

What each vendor actually meters

LangSmith (OFFICIAL, LangSmith pricing page, verified against third-party summaries current to August 2026). Developer is free with one seat and 5,000 free base traces a month. Plus is $39 per seat per month with 10,000 free base traces a month (shared across the account, not per seat, per the vendor's own FAQ). Beyond the free allowance, base traces (14-day retention) cost $2.50 per 1,000, and extended traces (400-day retention) cost $5.00 per 1,000 (an extra $2.50 per 1,000 on top of the base rate to upgrade). Enterprise is custom, billed annually. The meter is the trace itself: a trace with 2 spans and a trace with 20 spans are billed identically. Third-party guides that still cite $0.50 per 1,000 base traces are describing a retired price; the current rate is $2.50.

Braintrust (OFFICIAL, Braintrust pricing page). Starter is free: $10 of model credits, 1 GB of processed data then $4 per GB, 10,000 scores then $2.50 per 1,000, and 14-day retention. Pro is $249 a month, including $249 of model credits (recently raised from $100), 5 GB of processed data then $3 per GB, 50,000 scores then $1.50 per 1,000, and 30-day retention, extendable to 180 days at $0.50 per GB per month. Users are unlimited on every plan, so seats are not a cost driver. "Processed data" counts gigabytes across inputs, outputs, prompts, metadata, spans, datasets and attachments, so the bill follows payload size rather than trace count. Enterprise adds custom RBAC, SAML SSO, a BAA, SOC 2, self-hosting and custom retention, all at custom pricing (QUOTE-ONLY / UNKNOWN).

Arize AX (OFFICIAL, Arize documentation). AX Free: 25,000 spans and 1 GB of ingestion a month, 15-day retention. AX Pro is $50 a month with 50,000 spans and 10 GB of ingestion included, 30-day retention, and overages of $0.0008 per additional span and $3 per additional GB. One third-party cost tracker instead reported Pro overage at "$10 per million spans," which is $0.00001 per span and disagrees with Arize's own documented $0.0008 rate by a factor of 80; this article uses the documented $0.0008 rate. AX Enterprise is custom, with a reported typical range of $50,000 to $100,000 or more a year (REPORTED) and configurable retention, SOC 2, HIPAA and self-hosted deployment. Phoenix, Arize's open-source tracing layer, is free and unlimited to self-host.

Trace shapes and their telemetry footprint

The shapes are ILLUSTRATIVE: A, a simple chat request (2 spans, about 0.5 KB per trace); B, a RAG request with retrieval and reranking (5 spans, about 2 KB per trace); C, a multi-step agent with 10 or more tool and model calls (18 spans, about 8 KB per trace). LangSmith's meter ignores this column entirely; Arize's and Braintrust's do not.

Cost per 1 million traces by shape and volume

Assumptions: LangSmith seats scale with team size (1 seat at 100,000 traces a month, 3 at 1 million, 10 at 10 million, ILLUSTRATIVE); Braintrust scoring volume is 10% of traces (ILLUSTRATIVE); all figures use Braintrust Pro and Arize AX Pro extrapolated by their published overage rates, which is a reference calculation above the plan's intended scale, not a discount-negotiated enterprise price.

Formula: LangSmith = seats × $39 + max(0, traces − 10,000) ÷ 1,000 × $2.50; Braintrust = $249 + max(0, GB − 5) × $3 + max(0, scores − 50,000) ÷ 1,000 × $1.50, where GB = traces × KB per trace ÷ 1,000,000; Arize = $50 + max(0, spans − 50,000) × $0.0008 + max(0, GB − 10) × $3, where spans = traces × spans per trace.

Cost per 1 million traces, reading the table's per-million rate directly at 1 million volume: LangSmith is a flat $2,592 regardless of shape. Braintrust ranges from $324 to $333. Arize AX ranges from $1,610 to $14,410, an 8.9 times spread between the lightest and heaviest shape at the same trace count.

Sensitivity to spans per trace

Holding traces at 1 million a month, only Arize's bill moves with span count, because LangSmith counts traces and Braintrust counts gigabytes (which correlates with payload size, not span count, in this model):

Doubling spans per trace roughly doubles Arize's span-overage line, because each additional span is billed individually once the 50,000-span allowance is exhausted.

Sensitivity to retention

Retention changes the product tier required, not a separate line item, on LangSmith and Arize (extended retention is a different trace price on LangSmith; longer retention on Arize requires Enterprise). On Braintrust, retention beyond 30 days is a direct add-on: $0.50 per GB per month, billed for as long as the data is held, not a one-time fee. Formula: extended-retention cost = $0.50 × GB held × months held. At shape C's 8 GB per million traces, holding that data for six additional months costs a one-time total of $24 ($0.50 × 8 GB × 6 months); holding it for a full year costs a one-time total of $48 ($0.50 × 8 GB × 12 months). LangSmith's extended-trace upgrade of $2.50 per 1,000 traces on top of the base rate is a 100% surcharge for 400-day versus 14-day retention, applied per trace at write time, not retroactively.

Break-even: engineering and debugging hours saved

Formula: hours needed = monthly cost ÷ loaded hourly cost, with $85 an hour for a senior engineer (ILLUSTRATIVE). This is a threshold: the platform pays for itself if debugging time saved reaches this many hours, not a claim that it will.

At shape C volume, Arize's break-even threshold of 169.5 hours is roughly 4.2 standard 40-hour workweeks, or about one full-time engineer-month of saved debugging and root-cause time a month, which is a materially higher bar than LangSmith's flat trace-based bill at the same volume.

Budgeting traps

• Comparing a trace-metered bill to a span-metered bill without normalizing. A 2-span and an 18-span workload cost the same on LangSmith and up to 8.9 times apart on Arize.
• Ignoring payload size on Braintrust. A verbose agent that stuffs large tool outputs into context inflates "processed data" gigabytes even at a fixed trace count.
• Scoring volume. Braintrust's score meter is separate from its data meter; heavy automated evaluation can exceed the included score allowance well before the gigabyte allowance is exhausted.
• Extended retention as a per-write surcharge. LangSmith's extended-trace price applies going forward, not as a way to "rescue" old traces after the fact.
• Enterprise minimums. Both LangSmith and Arize AX Enterprise are custom-quoted, and Arize AX Enterprise was reported at $50,000 to $100,000 or more a year; this article's Pro-tier overage extrapolations above roughly 1 million traces a month should be treated as a reference ceiling, not an expected quote.

What to ask before you buy

Ask each vendor to price your actual trace shape: spans per trace, average payload size in kilobytes, and the share of traces you plan to score or evaluate. Then convert every quote to cost per 1 million traces at your real shape, not the vendor's headline number.

ARTICLE 2