LaunchDarkly’s additional usage units cover the products that were added to the feature flag platform: Experimentation, AgentControl (the product for configuring, evaluating and observing AI), observability and Data Export. The billing documentation states that depending on the billing model a customer may be billed for more than one of the account usage metrics on its Usage page.[1] The plans article covers prices; this article explains how each unit is counted and which legal terms attach.
Editions
On the pricing page, the Developer plan includes 100K experimentation MAU per month at no additional charge, 5K session replays and 5K errors, 10M logs and 10M traces, and 5K AI runs per month.[2] Foundation includes 5K AI runs and the same observability allowances, with usage beyond that billed as it occurs and AI runs at $5 per additional 1k.[2] Enterprise lists a custom number of AI runs and “Custom” volumes for session replays, logs, traces, errors and data export.[2] The billing page lets customers on other plan types add Experimentation, Data Export or observability to their plan, and the Data Export documentation says Data Export is an add-on feature to select plans, added through Sales.[6][4]
Experimentation
Depending on the billing model, a customer using Experimentation may be billed by Experimentation keys or by Experimentation MAU. If it is unsure which applies, the documentation tells it to contact its LaunchDarkly representative.[1]
Experimentation keys
Experimentation keys are the total number of unique context keys, from server-side, client-side, AI and edge SDKs, included in each experiment.[1] The same context key several times in one experiment counts as one Experimentation key. The same key in two different experiments counts as two.[1] The total measures how many contexts are targeted in experiments. The customer commits to a certain number of Experimentation keys each month, and usage over the entitlement is billed at the beginning of the next month. See Experimentation key.
Experimentation MAU
Experimentation MAU is the number of unique monthly active contexts, typically user contexts, evaluated by server-side, client-side, AI and edge SDKs and eligible to be included in an experiment.[1] It is based on the total number of unique contexts across the account for the month, whether or not they are actually in an experiment. Uniqueness is by context key, so changing a key counts as a new MAU, and a context in more than one experiment in a month still counts as one MAU.[1] The usage page for this model can be viewed by project, environment, SDK name, SDK app ID or anonymous contexts.[3] See Experimentation MAU.
The practical difference is breadth. Experimentation keys grow with the number of experiments that include the same users. Experimentation MAU grows with the whole eligible population, even if few are in experiments.
AgentControl and AI runs
AgentControl is billed by AI runs. An AI run is counted each time a model is called or a judge runs, whether standalone or attached to an AgentControl config.[1] That covers three scenarios: a model call in a production workflow, an online eval, or a direct judge call. AI runs are tracked automatically from events emitted by the AI SDK and recorded in the event recorder, and observability traces are optional and do not change the run count.[1] See AI run.
Online evaluations are more expensive. An online eval requires both a model run and a judge call, and these evaluations cost double because the action includes both. To reduce cost, the customer can lower the sampling percentage on its AgentControl config, which controls how often the judge runs on live traffic.[1] The model cost estimator on a config variation helps compare model costs but does not show LaunchDarkly billing.[1]
Plan terms differ. Developer includes 5,000 AI runs and ingestion stops at 5K. Foundation includes 5,000 and charges $5 per additional 1,000. Enterprise has a custom contracted number, and the comparison table lists data retention of 14, 30 and 100+ days respectively.[2] For annual plans with a commitment to AI runs, the true forward policy can adjust the commitment toward actual average use if the plan limit is consistently exceeded over a quarter.[1]
Observability
The observability feature is billed by four products: monthly ingested sessions, errors, traces and logs.[1] A customer is entitled to a certain number of each month, and LaunchDarkly bills at the beginning of each month for usage in excess of the prepaid entitlements from the prior month.[1] The pricing page defines the units. A log is a single log line and a trace is a single span, each ingested, stored and searchable for one month. A session replay is one recorded user session (a single tab instance), and an error is one unique error or exception, each stored for one month.[2] The documentation points to an observability usage calculator on the pricing page for estimates.[1] See Ingested session, Ingested error, Ingested trace and Ingested log.
Streaming Data Export
Customers that use streaming Data Export see a Streaming Data Export tab showing event usage, by event kind. When a customer subscribes to a plan it commits to a certain number of streaming Data Export events each month, and usage above the prepaid entitlement is billed at the beginning of each month for the prior month.[3] The documentation notes that a customer’s settings determine most of its Data Export volume.[4] See Streaming Data Export event and Data Export add-on.
Product-specific terms
Session Replay and AI Features carry extra terms that supplement the Agreement and prevail over it in a conflict. They are effective September 8, 2025.[5]
- Session Replay. The customer may use it only on sites and applications it has express permission to use or that it owns or controls. It must clearly disclose its use of Session Replay to relevant end users and, where data privacy laws require, obtain unambiguous, specific, freely given, informed and revocable consent from them.[5]
- AI Features. Input and Output are Customer Data and the customer owns them, but LaunchDarkly and its third-party providers own the AI Features. Third-party providers are prohibited from using Customer Data for their own model training, and the customer may opt out of AI Features through Admin settings.[5]
- AI restrictions. The customer may not use AI Features to develop its own AI products or machine learning models, to mislead anyone that an Output was solely human generated, to infringe rights, to violate the Acceptable Use Policy, to cause bias or discrimination, or to cause safety risks.[5]
See Session Replay and AI Features product-specific terms.
Practical points for asset managers
- Establish which Experimentation billing model each contract uses before comparing usage with entitlement.[1]
- Remember that an online eval costs double a single run, and that judge sampling rates affect cost directly.[1]
- Track the four observability categories separately, because each has its own entitlement.[1]
- Confirm who is allowed to enable Session Replay and AI Features, since each brings extra disclosure and use obligations.[5]
- Self-service terms charge excess usage in arrears monthly, so overage appears on the next invoice.[7]
Out of scope
Service connections and client-side MAU are in LaunchDarkly service connections and client-side MAU. Contract terms and support levels are in LaunchDarkly Terms of Service and support plans.