Langfuse vs LangSmith: Finding error records at month-end does not mean expired data has been deleted
When choosing a tracing platform for monthly reviews, compare access windows, actual deletion, evaluation-sample copies, and billing units separately. Langfuse’s historical access and LangSmith’s tiered retention support different management approaches; neither is universally the cheapest.
An application occasionally gives a wrong answer after launch, and the team wants to recover the original inputs, model outputs, and tool calls during its month-end review. When choosing Langfuse or LangSmith, the easiest mistake is not about which has a tracing dashboard, but about what will still be visible a month later and whether “not visible” means the data has been deleted. If you need a longer window for routine inspection and want several people to analyze records together, evaluate Langfuse Cloud first. If you want to manage many short-lived debugging records separately from a small number of longer-retained samples, LangSmith’s retention tiers are worth validating. Both require separate plans for handling sensitive fields and defining deletion scope. And includes no hands-on integration performance testing.
Langfuse currently makes an explicit distinction between data access and data retention. Hobby provides access to 30 days of history, Core to 90 days, and Pro and Enterprise to 3 years. The access window limits the history you can view; it does not automatically delete older data. For a team conducting monthly reviews, Core’s 90-day window is broader than one suited only to short-term debugging, but it does not justify promising customers that data disappears after 90 days. If automatic deletion is required, check the Data Retention feature and eligible plans separately.
LangSmith’s pricing page lists retention of 14 days for base traces and 180 days for extended traces. Its management documentation explains that expired traces can no longer be accessed through the interface or API. The associated user inputs and outputs are deleted from internal systems within the following day, although some analytics and billing metadata may remain. For teams that concentrate troubleshooting at month-end, 14 days may mean the evidence is already gone. Decide at collection time which records need extended retention rather than discovering the gap at month-end and trying to purchase it afterward. Changes to the current rules apply only to new records; existing records cannot all be assumed to follow the new settings.
The same month-end review requires different configurations
Suppose a team updates its prompts weekly and reviews poorly performing question-and-answer interactions at month-end. With Langfuse, it can first choose an access window covering the review cycle, then decide the project-level deletion period in days. With LangSmith, it needs to separate ordinary debugging records from those needed for longer-term comparisons. In the former case, guard against records expiring from the interface while remaining stored; in the latter, guard against original records expiring before the review starts. This is about matching work schedules to retention rules, not about either platform inherently meeting privacy requirements better.
LangSmith has another detail that can change both cost and retention: online evaluators or automation rules with retention extension enabled may upgrade matching base traces to extended traces. Current documentation says newly created evaluators and rules enable this extension by default, with an opt-out available during configuration. That does not mean every score, feedback item, or queue addition automatically raises the bill. Before configuring a quality check, also check whether it extends retention, then inspect the billing counts using a small sample.
In Langfuse, Data Retention is currently available on Cloud Pro, Enterprise, and the self-hosted Enterprise Edition. It is configured per project; you cannot assume that all projects in an organization inherit the same policy. Official documentation states that project retention does not delete audit logs or dataset items. Inputs, expected outputs, and metadata saved in datasets continue to exist. LangSmith datasets likewise have long-term retention behavior independent of traces. After a production record is copied into an evaluation sample, deleting the original trace therefore does not mean that every copy disappears. Teams should assign separate responsibility for dataset access and cleanup.
Different usage units: do not treat every 100,000 records as 100,000 conversations
Langfuse Core currently costs $29 per month and includes 100,000 billable units. The first overage tier costs $8 per 100,000 units, with unlimited users. Its units include data points such as traces, observations within workflows, and scores. Recording the same request in greater detail or adding scores can increase the number of units. LangSmith Plus costs $39 per seat per month, includes 10,000 base traces per month, and then charges for usage. Its traces can contain multiple steps in one application execution. For three members who require seats, the basic seat charge alone is $117 per month, before usage or other service charges.
These figures do not directly establish which is cheaper. Take the same batch of requests, count the actual data points, traces, extended-retention records, and people participating in analysis on each side, then obtain a complete quote. If automated evaluation uses models, check the associated inference charges too. Comparing only “100,000 units” with “10,000 traces” treats different units as the same capacity and overlooks the costs of changes in retention policy.
If your organization requires its own deployment, Langfuse officially provides a self-hosting route, but some additional features still require a license. Its architecture includes the application, background processing, and several storage components. Maintenance and backup responsibilities do not disappear from the cost comparison simply because cloud subscription prices are quoted. LangSmith’s pricing page places self-hosted and hybrid deployment in its custom Enterprise offering. Compare equivalent retention requirements, management capabilities, and operational services rather than an open-source deployment entry point with the price of a complete enterprise contract.
Finally, use an expiry drill to decide instead of looking only at feature lists: create sample traces containing no real customer information, copy one into a dataset item, configure the required retention policy, and check the behavior of the interface, API, datasets, and audit records separately. When several people need a shared, longer inspection window, start by calculating Langfuse’s unit-based billing and access windows. When the emphasis is on many short-term debugging records and a few extended records, validate LangSmith’s tiered retention first. The purchasing requirements should specify the review cycle, deletion scope, and accountable owners, not just say “supports log retention.”
Content note:This article is based on publicly available information and is intended as a reference for evaluating tools. Features, pricing, and related policies may change at any time; please refer to the latest official information.
Original source:langfuse.com