# Agent Pulse vs Langfuse

Langfuse and Agent Pulse both record model calls with their token usage and cost, and both can show cost per user. They are built for different jobs. Langfuse describes itself as an open-source AI engineering platform for debugging, analysing and iterating on AI agent applications. Agent Pulse records what each model and tool call costs and who it was for, and holds spend to budgets.

Everything said here about Langfuse is from its own documentation, linked beside each statement and listed at the end, as reviewed on 23 October 2026. Where this page interprets, it says so.

## What Langfuse is for

Langfuse's documentation describes three areas ([overview](https://langfuse.com/docs)):

- **Observability**: capturing every model call and other step in an application, multi-turn sessions, per-user tracking, dashboards and alerts.
- **Prompt management**: versioning prompts, testing them, and deploying them without code changes.
- **Evaluation**: LLM-as-a-judge, datasets, experiments, human annotation and custom scores.

It is open source and self-hostable ([overview](https://langfuse.com/docs)). Applications send it traces through its SDKs, wrappers around the OpenAI client, a LangChain callback handler, or OpenTelemetry ([get started](https://langfuse.com/docs/observability/get-started)).

## Where they overlap, and where they do not

Both answer what model calls cost. Langfuse takes usage and cost from the model's response, or infers cost from its own model definitions, to which you can add your own prices ([cost tracking](https://langfuse.com/docs/observability/features/token-and-cost-tracking)). Agent Pulse prices each call on the server from the token counts the provider reported, at the rate in force when the call was made ([how cost is calculated](/docs/how-cost-is-calculated)).

The difference is what each is for. Langfuse records the steps of an application so a team can debug, evaluate and improve it. Agent Pulse records counts, identifiers and timings, so spend can be read by your customer, user and request, and a budget can stop calls.

| | Agent Pulse | Langfuse |
| --- | --- | --- |
| Built for | The cost of each model and tool call, who it was for, and budgets ([AI agent metering](/ai-agent-metering)) | Debugging, analysing and iterating on AI applications ([overview](https://langfuse.com/docs)) |
| How it connects | Recorders inside your process, framework handlers, or OpenTelemetry spans; never in the path of model calls ([when Agent Pulse is unavailable](/docs/when-agent-pulse-is-unavailable)) | SDKs, OpenAI client wrappers, a LangChain callback handler, or OpenTelemetry ([get started](https://langfuse.com/docs/observability/get-started)) |
| How cost is worked out | Priced on the server from token counts, at the rate in force when the call was made ([how cost is calculated](/docs/how-cost-is-calculated)) | Ingested from the response, or inferred from model definitions; custom prices can be added ([cost tracking](https://langfuse.com/docs/observability/features/token-and-cost-tracking)) |
| Cost broken down by | Customer, user, request, agent, model and more ([cost per customer and user](/docs/cost-per-customer)) | Models, tags or users ([cost tracking](https://langfuse.com/docs/observability/features/token-and-cost-tracking), [users](https://langfuse.com/docs/observability/features/users)) |
| When spend runs high | Alerts as a budget nears its limit, and enforced budgets that stop covered calls ([budgets](/docs/budgets)) | Notifications when spend crosses a threshold ([cost tracking](https://langfuse.com/docs/observability/features/token-and-cost-tracking)) |
| Prompts and outputs | Never recorded: a record has no field for them ([what is recorded](/docs/what-is-recorded)) | Traces and observations carry inputs, outputs and metadata, which the SDKs can mask before sending ([masking](https://langfuse.com/docs/observability/features/masking)) |
| Prompt management and evaluation | Not offered | Prompt versioning and deployment, and evaluations ([overview](https://langfuse.com/docs)) |
| Where it runs | A service run by TechBridge Innovation ([privacy policy](/privacy)) | Open source and self-hostable ([overview](https://langfuse.com/docs)) |

## What the table leaves out

- **Debugging needs content; metering does not.** Langfuse can show what a step was given and what it returned, which is what debugging a prompt needs. Agent Pulse never sees prompts, completions, tool arguments or tool results, so it cannot help with that. It keeps a failure's error code, never its message ([what is recorded](/docs/what-is-recorded)).
- **Masking is configuration.** Langfuse's SDKs offer hooks that redact inputs, outputs and metadata before they are sent ([masking](https://langfuse.com/docs/observability/features/masking)). How much content reaches Langfuse depends on how a team sets it up. In Agent Pulse there is nothing to configure: content is never recorded.
- **Stopping a call.** An enforced Agent Pulse budget stops covered model calls at its limit. Whether it can stop a call before it is sent depends on how the call is recorded: the recorders for [LangChain and LangGraph](/docs/langchain), [Google ADK](/docs/google-adk), the [OpenAI Agents SDK](/docs/openai-agents), the [Vercel AI SDK](/docs/vercel-ai-sdk) and the [LiteLLM proxy](/docs/litellm) ask first. Calls recorded from OpenTelemetry spans count against budgets and raise alerts, but cannot be stopped, because the span arrives after the call.
- **Who a cost is for.** Both attribute cost to a user. Agent Pulse also files each call under one of your customers and a request, set once where your code checks the sign-in, so it can show what one customer or one request cost ([LLM cost tracking](/llm-cost-tracking)).

## Which should you choose?

Choose Langfuse when your team needs to see what each step of an application did, with its inputs and outputs, to debug it; when you want to version and deploy prompts or run evaluations; or when you need an open-source tool you can host yourself.

Choose Agent Pulse when the question is what your AI spend costs per customer, user and request; when you need budgets that can stop calls, not only report them; or when prompts and responses must not leave your process.

This is an interpretation of what each documents, not a measurement: the two answer different questions, and the right choice depends on which question you have. To see what Agent Pulse would show for your own code before connecting anything, [assess your codebase](/docs/assess).

## Sources

Langfuse's documentation, reviewed on 23 October 2026:

- [Langfuse overview](https://langfuse.com/docs)
- [Get started with observability](https://langfuse.com/docs/observability/get-started)
- [Token and cost tracking](https://langfuse.com/docs/observability/features/token-and-cost-tracking)
- [Users](https://langfuse.com/docs/observability/features/users)
- [Masking](https://langfuse.com/docs/observability/features/masking)