LangSmith is the observability product from LangChain, and the one most teams meet first because it is wired into LangChain and LangGraph by default. Its pricing has three parts that interact: a per-seat subscription, a per-trace usage charge, and a retention choice that changes the per-trace price. This post lays them out and works three monthly bills so you can place your own volume on the curve.
Prices are as published by LangChain in October 2026. LangChain changed its trace pricing in September 2026, so older write-ups quoting $0.50 per 1,000 base traces are out of date. Check the live pricing page before you sign anything.
The three plans
| Plan | Seats | Included traces | Price |
|---|---|---|---|
| Developer | 1 | 5,000 base traces a month | $0 |
| Plus | Unlimited | 10,000 base traces a month | $39 per seat per month |
| Enterprise | Custom | Custom | Custom, annual invoice |
Two things stand out. The included trace allowance does not scale with seats: a ten-seat Plus workspace gets the same 10,000 traces as a one-seat workspace, so the included volume is a rounding error for production and the real bill is the usage. And seats are billed in advance on the first of the month with no credit when you remove one, so a workspace that drifts to twelve seats pays for twelve.
What a trace is
One trace is one run tree: the top-level call and everything nested beneath it. An agent run that makes nine LLM calls, four tool calls and a retriever call is one trace. This matters when you compare with tools that meter per span or per observation, where the same run is fourteen billable units. Per run, LangSmith's unit is the generous one; the price per unit is what carries the bill.
Usage pricing: LSUs, base and extended
Usage beyond the included allowance is metered in LangChain Standard Units. One LSU is $1.00. The rates that matter:
| Item | Rate | Retention |
|---|---|---|
| Base trace | 0.005 LSU, so $5.00 per 1,000 | 14 days |
| Upgrade to extended | an additional $2.50 per 1,000 | up to 180 days |
| Extended trace, all in | $7.50 per 1,000 | up to 180 days |
For new traces since 14 September 2026 the extended retention ceiling is 180 days; traces created before that keep their original retention. The retention choice is per trace, so most teams keep base retention as the default and upgrade a sample, or the traces attached to an incident.
The LangChain pricing page itself lists only the LSU meter and "pay as you go"; the per-thousand figures above are the published LSU rates converted at $1 per LSU. If your account was quoted differently, your quote wins.
Three worked months
Assume a Plus workspace with three seats, every trace kept at base retention, and no Enterprise discount. Included traces are subtracted first.
| Monthly traces | Seats | Usage | Total |
|---|---|---|---|
| 1,000,000 | $117 | 990,000 traces at $5.00 per 1,000 = $4,950 | about $5,070 |
| 10,000,000 | $117 | 9,990,000 traces at $5.00 per 1,000 = $49,950 | about $50,070 |
| 50,000,000 | $117 | 49,990,000 traces at $5.00 per 1,000 = $249,950 | about $250,070 |
Upgrading everything to extended retention multiplies the usage line by 1.5. Upgrading 10% of traces adds $2.50 per 1,000 on that slice: at 1M traces a month, another $250.
A million traces a month is roughly 33,000 runs a day, which is a mid-sized production agent, not a large one. At that volume the seat line is 2% of the bill. The pricing is designed for the development loop, where a team of ten runs a few hundred evaluations a day and the seats are the bill, not for a fleet of production agents where every run is a trace.
Where the bill surprises people
- The allowance is per workspace, not per seat. Adding seats raises the subscription without raising the included traces.
- Every production run is a trace. Teams that turned on tracing in development and shipped with it on discover the usage line a month later. Sampling in production is the usual fix, and it costs you the runs you did not sample when something goes wrong.
- Retention is short by default. Fourteen days means the trace from the run a customer is complaining about has often expired by the time the ticket reaches an engineer. Keeping extended retention on everything is the 1.5x multiplier above.
- Seats bill in advance and do not refund. Audit the seat count before the first of the month.
What this buys you
LangSmith's strengths are real: the tightest LangChain and LangGraph integration, a strong evaluation and dataset workflow, prompt management, and a playground that lets a team iterate on a chain quickly. If your team lives in LangGraph and spends most of its time in the development loop, the per-seat part of the bill is fair value and the trace volume stays small.
Where it fits less well is the production side: thousands of runs a day, each one a trace, kept long enough to debug a customer complaint, with the cost of each run attributed to a customer. That is the shape of bill where $5 per thousand compounds, and it is the case Agent Ping is built for: a flat monthly tier with named limits, no per-seat charge, and cost per customer and feature on every run. The Agent Ping vs LangSmith comparison sets the two side by side, feature by feature.
If you are working out a budget, the LLM token cost calculator estimates what the model calls inside each run cost, which is usually the larger number and the one observability is supposed to help you control.