NewIn beta. Read the docs

AI observability: cost, tokens and latency for every model call

Wrap the Vercel AI SDK, the OpenAI client or the Anthropic client once, or send the events yourself, and every generation shows up with its cost, tied to the user and the session behind it.

Start free trialSee live demo
1
packageWraps the Vercel AI SDK, the OpenAI client and the Anthropic client. Every provider they reach, Azure and OpenRouter included.
4
event kindsgeneration, embedding, span and trace, as ordinary events with flat ai_* properties.
30
days of contentPrompts and completions are kept that long, if you store them at all. Cost and tokens stay.
Cost

What a user costs you

Every call carries its model, tokens and cost, and the profile, session and group it belongs to. So the question is not what the API bill was, but which feature, which plan and which customer it went to.

A price catalog, or your own cost
Major providers are priced for you. Send cost when a gateway already knows it.
Per model, feature, user and group
The AI overview breaks cost down by all four.
On the profile and the session
An AI usage widget with lifetime cost, tokens, latency and errors.
Read the docs
Traces

Traces for agents

An agent runs the model several times and calls tools in between. Wrap the request once and every step lands in one trace, laid out in time, with the step that failed in red.

Multi-step chains and tool calls
withAiContext groups a request; nested contexts merge over it.
Errors and latency per step
Time to first token on streams, the finish reason, the error that was rethrown.
Conversations across requests
Pass a conversationId and the thread reads as one.
Privacy

Prompts only if you want them

Tokens, cost, latency and errors are analytics. Prompts and completions are content. OpenPanel keeps the first always and the second only when you say so, and never for more than 30 days.

Content storage is a switch
Off for the project, or privacyMode per call.
30-day expiry
Stored content is dropped after a month. The numbers stay with your other events.
Self-host for full control
The events land in your own ClickHouse.

In and out

How the events get in, and where else they show up.

How it works

Three steps, then it is just events

  1. 1

    Wrap the model

    One middleware on wrapLanguageModel, with the OpenPanel client you already have.

    import { wrapLanguageModel } from 'ai';
    import { openai } from '@ai-sdk/openai';
    import { openpanelAiMiddleware }
      from '@openpanel/ai/vercel';
     
    const model = wrapLanguageModel({
      model: openai('gpt-5'),
      middleware:
        openpanelAiMiddleware({ op }),
    });
  2. 2

    Say who and what

    profileId, a feature name and groups per call, or once per request with withAiContext.

    await generateText({
      model, messages,
      providerOptions: { openpanel: {
        profileId: user.id,
        name: 'inbox-summary',
        groups: [user.orgId],
      } },
    });
  3. 3

    Read cost, traces and the people behind them

    The AI section, the profile, the session, and any report you build on ai_* properties.

Works together with

Frequently asked questions

Common questions about AI observability in OpenPanel.

Start understanding your users today

One script on your site, or one SDK call in your app, and the first events show up within a minute. 30-day free trial, no credit card.

Start tracking your AI callsSee pricing

Cloud from $2.50 a month, or self-host for free.