Tokens are the atoms of LLM costs. Every prompt you send and every response you receive is metered in tokens, and the price difference between models can be 100x. A real-time token usage dashboard isn't a nice-to-have — it's essential.
What AgentBurn Tracks
Every cost event ingested into AgentBurn includes input tokens, output tokens, model name, and dollar cost. The dashboard aggregates this into actionable views:
- Cost over time — Line chart showing daily/weekly/monthly spend with trend lines
- Provider breakdown — Pie chart showing OpenAI vs Anthropic vs others as a percentage of total spend
- Agent leaderboard — Which agents are your biggest spenders?
- Model comparison — Are you overpaying by using GPT-4o where GPT-4o-mini would suffice?
- Token efficiency — Input-to-output token ratio reveals prompt bloat
Reading the Dashboard
The most important metric isn't total spend — it's cost per task. If your customer support agent costs $0.12 per ticket resolution, that's great. If it costs $2.40, you have a prompt engineering problem.
AgentBurn's task and workflow grouping lets you attribute costs to business outcomes, not just API calls. Tag your cost events with taskId and workflowId to unlock this view.
Spotting Anomalies
The time-series chart is your early warning system. A sudden spike means something changed — a new prompt, a retry loop, or a traffic surge. Combined with budget alerts, you'll catch issues within minutes, not days.
Input vs Output Token Ratio
A healthy agent typically has a 2:1 to 5:1 input-to-output ratio. If you see 20:1, your system prompts are bloated. If you see 1:5, your agent is generating too much text. Both indicate optimization opportunities.