---
type: use-case
title: LLM cost dashboard — see spend by project, agent, model, and route
description: A practical LLM cost dashboard for teams that need token spend, latency, errors, budget alerts, and model breakdowns before the provider invoice arrives.
last_updated: 2026-05-09
audience: AI founders / platform engineers
---

# LLM cost dashboard

Provider invoices answer one question: how much did the account spend? They usually do not answer the operational questions an engineering team needs during the month.

Sutrace answers:

- Which project is spending the most?
- Which agent crossed its daily cap?
- Which route became expensive after a deploy?
- Which model has the highest p95 latency?
- Which failed calls still consumed tokens?
- Which customer workflow caused the spike?

## What the dashboard shows

The Agents dashboard rolls up every SDK, proxy, and local coding-agent event into one workspace view:

| View | What it shows |
|---|---|
| Calls | Recent provider calls with status, model, tokens, cost, latency, and route |
| Spend | Cost over time and total cost in the current window |
| Latency | p50 and p95 latency across provider calls |
| Errors | Error rate, failed calls, and provider status |
| Project breakdown | Spend and calls by project |
| Agent breakdown | Spend and calls by agent |
| Model breakdown | Spend and calls by model |

## Install path

```ts
import OpenAI from "openai";
import { wrapOpenAI } from "@sutrace/llm";

const openai = wrapOpenAI(new OpenAI(), {
  apiKey: process.env.SUTRACE_API_KEY,
  project: "support-agent",
  agent: "refund-router",
  route: "/tickets/summarise",
  budget: { perRunUsd: 1, dailyUsd: 20, monthlyUsd: 500 },
});
```

Smoke test:

```bash
npx @sutrace/llm test --api-key "$SUTRACE_API_KEY"
```

## Why this sells first

It has a short path to value. A team does not need to replace Datadog, LangSmith, Grafana, or Helicone to try it. They wrap one call, send one event, and see the first useful signal.

## What Sutrace does not need

For the first cost dashboard, Sutrace does not need prompt or completion text. Metadata is enough:

- provider
- model
- input tokens
- output tokens
- status
- latency
- project
- agent
- route

That makes the first install easier for teams with privacy or compliance concerns.