Direct answer: for the same production workload — 1,000,000 LLM requests a month, 6 observations per trace, 10% sampled evaluation — official list prices on 2 November 2026 give roughly USD 5,340/month on LangSmith Plus (10 seats plus 990,000 base traces), USD 528/month on Langfuse Cloud Core (7.1 million units), USD 333/month on Braintrust Pro (8 GB processed data plus 100,000 scores), while Arize AX only quotes Enterprise once you exceed 50,000 spans. The spread is about 16x, not because one vendor is “cheaper”, but because the four meter four different things: LangSmith counts traces, Langfuse adds traces, observations and scores, Braintrust counts ingested bytes, and Arize counts spans. Break your own traffic into those four units before comparing unit prices.
This guide is for engineering leaders, platform teams and procurement groups choosing an LLM observability or evaluation platform. Data cut-off: 2 November 2026; prices come from the official LangChain, Langfuse, Braintrust and Arize pricing pages and documentation, in USD, excluding tax, with no FX conversion. The 1-million-request model is an editorial calculation, not a quote; trace length, observation count, scoring rate and retention must be replaced with your own data.
Ask “what is counted” before “how much”; the same trace can be 1 trace, 7.1 units, 8 KB or 6 spans; retention, evaluation and seats are the three add-ons most often missed.
What is the difference between trace, span, observation and unit?
A trace is one complete request; an observation is one step inside it (a model call, a tool, a retrieval); a score is one evaluation result; a span is how Arize counts steps inside a trace; Langfuse’s unit is the sum of traces, observations and scores. The same trace can be 1 unit at LangSmith, 7.1 units at Langfuse, 6 spans at Arize, or a handful of bytes at Braintrust.
“Observability platform” sounds like one product category, but the billing objects are different, so the “from USD 29/month” on a homepage cannot be compared directly. Align the four terms first:
| Platform | What is counted | One “standard trace” counts as | Direct effect |
|---|---|---|---|
| LangSmith | Base trace | 1 trace = 1; model calls and tool steps inside it are not billed separately | Heavier traces are relatively cheaper; longer retention is charged per trace |
| Langfuse Cloud | Unit = trace + observation + score | 1 trace + 6 observations + 0.1 score ≈ 7.1 units | More auto-instrumentation and more scores mean more units |
| Braintrust | Processed data (GB) + score | Billed by ingested bytes; 8 KB ≈ 0.000008 GB | Longer traces and attachments cost more; deleting does not reduce the meter |
| Arize AX | Span + ingestion volume (GB) | 6 steps in one trace = 6 spans | More steps cost more; above 50,000 spans requires Enterprise |
Sources: Langfuse Billable Units (Units = Count of Traces + Count of Observations + Count of Scores), LangChain Pricing FAQ (a trace can contain many steps), Braintrust Plans and limits (processed data includes traces and spans, measured at ingestion), Arize Pricing (trace spans and ingestion volume listed separately). Accessed 2 November 2026.
That is why a “cost per 1,000 requests” can differ by an order of magnitude: LangSmith sells a per-request flat fee, Langfuse sells data points, Braintrust sells bytes, and Arize sells steps. Reading only the entry price leads to the wrong conclusion.
What are the official 2026 list prices?
LangSmith Developer is free with 5,000 base traces/month; Plus is USD 39/seat with 10,000 included, then USD 0.005/base trace. Langfuse Hobby is free with 50,000 units; Core is USD 29 with 100,000 included and USD 8/100,000 after; Pro is USD 199 and Enterprise USD 2,499. Braintrust Starter is free with 1 GB plus 10,000 scores; Pro is USD 249 with 5 GB plus 50,000 scores. Arize AX Free includes 25,000 spans; Pro is USD 50 for 50,000 spans, and beyond that is an Enterprise quote.
| Platform | Plan | Monthly fee | Included | Overage / add-on | Retention |
|---|---|---|---|---|---|
| LangSmith | Developer | USD 0 | 5,000 base traces; 1 seat | — | 14 days |
| LangSmith | Plus | USD 39/seat | 10,000 base traces | USD 0.005/trace; extended +0.0025 | 14 days; 180 days extended |
| LangSmith | Enterprise | Custom | Custom | Custom; self-hosted or hybrid | Custom |
| Langfuse | Hobby | USD 0 | 50,000 units; 2 users | — | 30 days |
| Langfuse | Core | USD 29 | 100,000 units | USD 8/100,000 units, graduated | 90 days |
| Langfuse | Pro | USD 199 | 100,000 units | Same; Teams add-on +USD 300 | 3 years |
| Langfuse | Enterprise | USD 2,499 | 100,000 units | Same; annual volume pricing | 3 years |
| Braintrust | Starter | USD 0 | 1 GB + 10,000 scores | USD 4/GB; USD 2.50/1,000 scores | 14 days |
| Braintrust | Pro | USD 249 | 5 GB + 50,000 scores | USD 3/GB; USD 1.50/1,000 scores | 30 days; +USD 0.50/GB/month up to 180 days |
| Braintrust | Enterprise | Custom | Custom | Custom; BYOC or self-hosted | Up to 365 days |
| Arize AX | Free | USD 0 | 25,000 spans + 1 GB | — | 15 days |
| Arize AX | Pro | USD 50 | 50,000 spans + 10 GB | Not published | 30 days |
| Arize AX | Enterprise | Custom | Custom | Custom | Custom |
Sources: LangChain Pricing, Langfuse Pricing, Braintrust Pricing and Plans and limits, Arize Pricing (all accessed 2 November 2026). USD, excluding tax; LangSmith’s LSU is USD 1; Braintrust’s GB is measured at ingestion (this model assumes 1 GB = 1,000 MB); Arize’s overage is not public, so the table says “not published” instead of guessing. Langfuse’s graduated rates are USD 8 per 100,000 for 100k–1M, 7 for 1M–10M, 6.50 for 10M–50M and 6 above 50M.
What does 1 million requests cost per month?
With 1,000,000 traces, 6 observations each, 10% scored, 8 KB average, 10 seats and 30-day retention: about USD 5,340 on LangSmith, USD 528 on Langfuse Core, USD 333 on Braintrust Pro, and an Enterprise quote on Arize.
Write the assumptions down first, or the comparison cannot be reproduced:
- 1,000,000 traces/month, 6 observations per trace (1 entry plus 5 model or tool calls).
- 10% of traces get one score each, for 100,000 scores.
- 8 KB per trace on average (inputs, outputs and metadata), or 8 GB/month.
- 10 engineering seats; 30-day retention; excluding tax and judge-model tokens.
| Platform | Calculation | Monthly fee (USD) | Per 1,000 requests |
|---|---|---|---|
| LangSmith Plus | 10 × 39 + (1,000,000 − 10,000) × 0.005 | 5,340 | 5.34 |
| Langfuse Core | 29 + 900,000 ÷ 100,000 × 8 + 6,100,000 ÷ 100,000 × 7 | 528 | 0.53 |
| Braintrust Pro | 249 + (8 − 5) × 3 + (100,000 − 50,000) ÷ 1,000 × 1.50 | 333 | 0.33 |
| Arize AX | 6,000,000 spans > 50,000; overage not published | Enterprise quote | Unknown |
Editorial calculation, not a quote. LangSmith is priced at 10 seats; Langfuse units = 1M traces + 6M observations + 100k scores = 7.1M; Braintrust includes 5 of the 8 GB in Pro and bills 3 GB at USD 3; 50,000 of the 100,000 scores are included and 50,000 bill at USD 1.50/1,000.
Unpack the formulas: of LangSmith’s USD 5,340, 4,950 is trace usage and 390 is seats; of Langfuse’s USD 528, 29 is the platform fee and 499 is unit overage; of Braintrust’s USD 333, 249 is the platform fee, 9 is 3 GB of data and 75 is 50,000 extra scores. The platform fee is about 7%, 5% and 75% of the total respectively — which explains who is more expensive at low volume and who is cheaper at high volume.
Evaluation models are not included yet. If the 100,000 scores come from an LLM-as-judge, the judge tokens are paid to the model provider on all four platforms; the platform-side score charge is separate, as the next section shows.
Why is LangSmith nearly 10x Langfuse?
Because the billing object differs: LangSmith charges USD 0.005 per trace and does not charge for the 6 steps inside it; Langfuse turns the trace, 6 observations and score into units, but each unit costs about USD 0.00007–0.00008. The unit price differs by more than 60x, the counted volume by roughly 7x, so the bill differs by about 10x.
In effective cost per trace: LangSmith about USD 0.00534, Langfuse about USD 0.000528, Braintrust about USD 0.000333. That does not make LangSmith unreasonable — it offers a predictable flat fee per trace, which suits teams with heavy traces, many steps and no appetite for per-span billing. At the other end, light traces with few steps make per-trace billing expensive.
The real question is: how many steps and kilobytes does my average trace have, and how long must I keep it — not “who is cheaper per 1,000 requests”. At 1 million requests with 20 steps per trace, Langfuse’s count rises to 21 million units while LangSmith still counts 1 million traces.
Do the rankings flip when traces get lighter or heavier?
Yes. With light traces (2 observations, 2 KB), Langfuse is about USD 248 and Braintrust about USD 324; with heavy traces (20 observations, 200 KB), Langfuse is about USD 1,453 and Braintrust about USD 909. LangSmith stays near USD 5,340 because it bills per trace, not per volume.
| Workload (all 1M traces, 10% scored) | LangSmith | Langfuse Core | Braintrust Pro | Lowest |
|---|---|---|---|---|
| Light: 2 observations, 2 KB | 5,340 | 248 | 324 | Langfuse |
| Medium: 6 observations, 8 KB | 5,340 | 528 | 333 | Braintrust |
| Heavy: 20 observations, 200 KB | 5,340 | 1,453 | 909 | Braintrust |
USD/month. Light Langfuse units = 1M + 2M + 100k = 3.1M; heavy units = 1M + 20M + 100k = 21.1M, which enters the 10M–50M tier. Braintrust’s light 2 GB sits inside the 5 GB allowance; the heavy 200 GB bills 195 GB at USD 3/GB.
Braintrust overtakes Langfuse on heavy traces because it bills bytes rather than data points, while Langfuse grows linearly with observations. For very light traces at very high volume, Langfuse’s unit price wins. For Braintrust’s GB charge to exceed LangSmith’s per-trace charge in this model (10% scoring, 10 seats), the average trace would need to exceed about 1.6 MB — rare for a normal agent trace.
This is also why Arize cannot be priced from its public page: Free and Pro cover only 25,000 and 50,000 spans, and anything beyond is Enterprise. At 6 million spans, the public page gives no number, and any “price per span” would be a guess.
Retention, evaluation and seats: the three most-missed add-ons
LangSmith keeps base traces for only 14 days; extending to 180 days adds USD 0.0025 per trace. Langfuse builds 30 days, 90 days and 3 years into its plans. Braintrust adds USD 0.50/GB/month after 30 days. Arize covers 15 or 30 days, with longer retention on Enterprise. On evaluation, LangSmith’s Tuned Evaluator costs 0.015 LSU per successful run, Langfuse counts each score as one unit, Braintrust bills each score, and Arize includes unlimited evals in the plan.
| Item | LangSmith | Langfuse | Braintrust | Arize AX |
|---|---|---|---|---|
| Base retention | 14 days | 30 / 90 days / 3 years | 14 / 30 days | 15 / 30 days |
| Extension | 180-day extended, +USD 0.0025/trace | Upgrade the plan | +USD 0.50/GB/month, up to 180 days | Enterprise |
| Cost for 1M traces | About USD 7,425/month usage (+2,475) | Pro from USD 199 includes 3 years | 8 GB about +USD 4/month; 200 GB about +USD 100/month | Quote |
| Evaluation billing | Tuned Evaluator 0.015 LSU/run | 1 unit per score | Per score | Unlimited in plan |
| 100k evaluations, incremental | About USD 1,500 (+250 upgrade) | About USD 7 (marginal) | About USD 75 | 0 (platform side) |
“100k evaluations, incremental” is sized to this model: LangSmith 100,000 × 0.015 = USD 1,500, plus 100,000 × 0.0025 = USD 250 because scored traces are automatically upgraded to extended; Langfuse adds 100,000 units in the 1M–10M tier, about USD 7; Braintrust Pro includes 50,000 scores, so 50,000 × USD 1.50/1,000 = USD 75. Judge-model tokens are excluded. LangSmith’s Tuned Evaluator is in public beta in the US cloud; availability is per contract.
Write retention as a specific number of days, not “as needed”. LangSmith’s 14-day base limit is hard, with per-trace upgrades beyond it; Braintrust meters processed data at ingestion, so deleting data does not refund the meter; Langfuse’s “data access” is how far back you can look, which is different from “data retention” (permanent deletion) and is configurable on Pro and above.
Seats deserve their own line: LangSmith Plus is USD 39/seat, so 10 seats is USD 390; Langfuse Core and above have unlimited users, but SSO/RBAC requires Pro plus the USD 300 Teams add-on; Braintrust’s SAML SSO is Enterprise-only; Arize has unlimited users on every tier, but SSO and audit logs are Enterprise. Put “10 seats plus SSO” in the same table and the conclusion may differ from platform fee alone.
Is self-hosted Langfuse or Phoenix really free?
The software licence is free: Langfuse OSS is MIT and Arize Phoenix can be self-hosted locally. But servers, Postgres/ClickHouse/Redis/object storage, backup, upgrades, security and staff time are yours. Below 50,000 requests a month, the free SaaS tiers are usually cheaper; above 1 million with data that cannot leave your VPC, self-hosting or Enterprise BYOC makes more sense.
| Self-hosting cost | Question to answer |
|---|---|
| Compute and storage | What is peak ingest in GB/day? How much disk for 90-day retention? |
| Dependencies | Who runs Postgres, ClickHouse, Redis and S3? Who upgrades them? |
| Backup and DR | What are RPO/RTO? Is it cross-AZ? |
| Security and compliance | Who can see traces? Is PII included? Where are audit logs? |
| Staff time | How many hours a month for on-call, upgrades and incidents? |
| Opportunity cost | Should those hours go to the product instead? |
Langfuse’s documentation states that self-hosted OSS is free under the MIT licence with no usage-based billing; Arize Phoenix is open-source and self-hostable. Server and staff prices vary by company, so this article gives no universal number rather than mislabel “free software” as “zero cost”.
The biggest advantage of self-hosting is not “zero licence fee” but the data boundary and customisation; the biggest risk is swapping a predictable SaaS fee for an open-ended operations project. A rough “requests per month × operations minutes per trace × staff cost” usually lands closer to reality than the word “free”.
What should procurement and acceptance specify?
At least seven things: the billing unit, whether failures and retries are billed, how many observations auto-instrumentation creates by default, retention and deletion, evaluation and judge-model costs, seats/SSO/RBAC, and data region plus export. Run a two-week POC on production traffic to measure units/GB/spans per 1,000 requests, then extrapolate.
| Acceptance item | Ask the vendor | Evidence to obtain |
|---|---|---|
| Billing unit | How are trace, span, observation, score and GB counted? | Billing doc plus a recent invoice sample |
| Failures and retries | Are timeouts, 429s and duplicate ingestion billed? | Written rule plus POC invoice |
| Auto-instrumentation | How many spans does the SDK create by default? Can we sample? | Default config plus sampled unit count |
| Retention and deletion | What do 30/90/180 days cost? Does deletion affect billed usage? | Price list plus deletion policy |
| Evaluation | Are LLM-as-judge scores billed? Who pays judge tokens? | Eval price list plus model bill |
| Seats and permissions | Are SSO, RBAC and audit logs extra? | Plan comparison plus contract terms |
| Data and export | Which region stores the data? Can we export everything? | DPA plus export API docs |
| Overage and price lock | Do graduated tiers apply automatically? Can an annual contract lock the price? | Quote plus renewal terms |
This is a procurement checklist, not any vendor’s contract terms. Prices and features are governed by the official quote and contract at signing.
Which option fits which scenario?
Start a POC on the free tiers; for cost-sensitive workloads with many steps, prefer Braintrust Pro or self-hosted Langfuse; for 3-year retention plus SSO and audit, use Langfuse Pro+Teams or Enterprise; heavy LangChain/LangGraph users get the best experience from LangSmith, but per-trace billing is the most expensive at volume; if data cannot leave the VPC, self-host or use Enterprise BYOC.
| Scenario | First choice | Why | Watch out |
|---|---|---|---|
| Individual POC, small team | LangSmith Developer / Langfuse Hobby / Braintrust Starter / Arize Free | Free allowance is enough to validate | Retention is only 14–30 days |
| Cost-sensitive, many steps per trace | Braintrust Pro or self-hosted Langfuse | Per GB or per unit is cheaper for heavy traces | Self-hosting needs staff time |
| 3-year retention, SSO and audit | Langfuse Pro+Teams / Enterprise | Long retention and permissions are in the plan | Teams adds USD 300/month |
| Deep LangChain / LangGraph user | LangSmith | Best integration and debugging | Per-trace billing is the most expensive at volume |
| Data cannot leave the VPC | Self-hosted Langfuse / Phoenix, or Braintrust Enterprise BYOC | Data plane stays in your environment | Operations and upgrade cost |
| Consolidate into existing APM | Price the LLM workload separately first | Avoid hiding it behind an APM discount | APM pricing was not verified here |
“First choice” is an editorial judgement, not a vendor recommendation; the final decision depends on data boundary, retention, permissions and the existing stack.
FAQ
What is the difference between a trace, a span, an observation and a unit?
A trace is one complete request; an observation is one step inside it (a model call, a tool, a retrieval); a span is how Arize counts steps inside a trace; Langfuse’s unit is traces plus observations plus scores. LangSmith bills only traces, so a trace with 6 steps still counts as 1; Langfuse counts 7 units (1 trace + 6 observations), plus any score.
What does 1 million requests cost per month?
Under this article’s assumptions (6 observations per trace, 10% scored, 8 KB, 10 seats, 30-day retention): about USD 5,340 on LangSmith Plus, USD 528 on Langfuse Core and USD 333 on Braintrust Pro. Arize requires an Enterprise quote above 50,000 spans. Your trace length, step count and scoring rate will change the result.
Why is LangSmith so much more expensive than Langfuse?
Because the billing object differs. LangSmith charges USD 0.005 per trace and does not bill the steps inside it; Langfuse counts the trace, every observation and every score as a unit, but each unit costs about USD 0.00007–0.00008. The unit price differs by about 60x and the volume by about 7x, so the bill differs by about 10x. The heavier the trace, the more the per-trace flat fee works in LangSmith’s favour.
Are LLM-as-judge evaluations billed separately?
Separate the platform side from the model side. Platform: LangSmith’s Tuned Evaluator costs 0.015 LSU (about USD 0.015) per successful run, with failures and skips not billed; Langfuse counts each score as one unit; Braintrust bills each score (USD 1.50/1,000 on Pro); Arize includes unlimited evals in the plan. Model: judge tokens are paid to the model provider on all four platforms.
How much does 30-day versus 3-year retention cost?
LangSmith base traces include 14 days; extending to 180 days adds USD 0.0025 per trace, and upgrading all 1 million traces takes usage to about USD 7,425/month, about USD 2,475 more than base. Langfuse puts 90 days in Core and 3 years in Pro (from USD 199). Braintrust adds USD 0.50/GB/month after 30 days: about USD 4/month for 8 GB and USD 100/month for 200 GB. Arize covers 15/30 days, with longer retention on Enterprise.
Is self-hosted Langfuse or Phoenix free?
The software licence is free (Langfuse OSS is MIT; Phoenix is self-hostable), but servers, databases, object storage, backup, upgrades, security and staff time are yours. Below 50,000 requests a month, the free SaaS tiers are usually cheaper; above 1 million with data that cannot leave the VPC, self-hosting or Enterprise BYOC makes more sense.
Why can’t Arize give a unit price?
Because Arize AX Free and Pro publish only 25,000/50,000 included spans and 1/10 GB, with no public overage rate — beyond that it is Enterprise custom. At 6 million spans the public page cannot produce a monthly fee; ask for a quote and an invoice sample sized to your spans, ingest GB and retention days.
How do I estimate my own bill from real data?
Run two weeks of production traffic and record four numbers: traces, observations per trace, average bytes and scoring rate. Then apply: LangSmith = seats × 39 + max(0, traces − 10,000) × 0.005; Langfuse = plan fee + graduated unit fees; Braintrust = plan fee + excess GB × rate + excess scores × rate. Calibrate on 1,000 real traces before scaling to the month; it beats reading the entry price.
Research method and limitations
This article separates “official facts” from “editorial modelling”: monthly fees, included volume, overage rates, retention days and unit definitions come from official pricing pages or documentation; the 1-million-request bill, light/medium/heavy comparison, cost per 1,000 requests and evaluation increments are arithmetic on a single set of assumptions, not a vendor quote.
Five limitations: (1) only public list prices are compared, without enterprise discounts, annual commitments or cloud-marketplace private offers; (2) amounts are in USD with no FX conversion; (3) 8 KB per trace, 6 observations and 10% scoring are assumptions that can differ several-fold in practice; (4) Arize’s overage is not public, so no like-for-like monthly fee can be given; (5) judge-model tokens, staff time, migration and training are excluded from TCO. Treat this as a unit-alignment template and replace the assumptions with two weeks of your own traffic.
Sources and basis
- LangChain, Pricing: langchain.com/pricing. Developer USD 0 with 5,000 base traces and 1 seat; Plus USD 39/seat with 10,000 included; USD 0.005 per base trace and USD 0.0025 for an extended upgrade; 1 LSU = USD 1; Tuned Evaluator 0.015 LSU/run; 14-day base and 180-day extended retention; Enterprise custom. Accessed 2 November 2026.
- LangChain, Pricing FAQ: langchain.com/pricing. A trace is one application execution and can contain many steps; Tuned Evaluators bill only successful runs, not failures or skips; scored traces are automatically upgraded to extended.
- Langfuse, Pricing: langfuse.com/pricing. Hobby 50k units; Core USD 29/100k; Pro USD 199; Enterprise USD 2,499; overage from USD 8/100k with 7/6.50/6 tiers; Pro includes 3-year access and Teams +USD 300; Enterprise includes audit logs, SCIM and an SLA. Accessed 2 November 2026.
- Langfuse, Billable Units: langfuse.com/docs/administration/billable-units. Units = Count of Traces + Count of Observations + Count of Scores; self-hosted OSS is free under MIT with no usage-based billing.
- Braintrust, Pricing: braintrust.dev/pricing. Starter USD 0 with 1 GB + 10k scores; Pro USD 249 with 5 GB + 50k scores; USD 3/GB and USD 1.50/1,000 scores after; USD 0.50/GB/month after 30 days up to 180 days; Enterprise custom. Accessed 2 November 2026.
- Braintrust, Plans and limits: braintrust.dev/docs/plans-and-limits. Processed data is the ingested bytes of logs, experiments and datasets, including inputs, outputs, prompts, metadata, traces and spans, measured at ingestion so deletion does not reduce it; scores are counted per record; Starter 14 days, Pro 30 days. Accessed 2 November 2026.
- Arize, Pricing: arize.com/pricing. AX Free 25k spans + 1 GB, 15 days; AX Pro USD 50 for 50k spans + 10 GB, 30 days; Enterprise custom; users and evals unlimited on every tier. Accessed 2 November 2026.
Data cut-off 2 November 2026; prices are public list prices excluding tax; all models are editorial calculations. This is not procurement, legal or financial advice.