Data
A worked example of a month of multi-provider LLM spend — the breakdowns that matter, the anomalies worth catching, and how project-level attribution changes the conversation.
This post walks through a realistic example month of AI spend for a mid-size engineering team, the way it looks in Model Meter's dashboard. The numbers below are illustrative — they represent a plausible month for a team of ~15 engineers shipping AI features across four projects. Your data will differ. The point is what the breakdown makes possible.
Total tracked AI spend for the month: $12,544.
| Provider | Spend | Share | Top model |
|---|---|---|---|
| OpenAI | $5,890 | 47% | gpt-4o |
| Anthropic | $3,756 | 30% | claude-3-5-sonnet |
| $1,879 | 15% | gemini-2.0-flash | |
| Groq | $1,019 | 8% | llama-3-70b |
This is the kind of distribution that surprises teams. Anthropic often gets assumed to dominate because Claude handles the harder reasoning tasks. Here, OpenAI leads because gpt-4o-mini carries the highest call volume across high-frequency features.
| Project | Spend | Primary driver |
|---|---|---|
| API platform | $4,820 | gpt-4o for structured extraction |
| Engineering copilot | $3,170 | claude-3-5-sonnet for code tasks |
| Support intelligence | $2,984 | gpt-4o-mini for ticket classification |
| R&D experiments | $1,570 | Mixed — Gemini Flash + Groq |
The engineering copilot is the interesting line here. New features ramp quietly, and by the time the invoice arrives they're already the second-largest cost driver. Seeing that mid-month is what lets you set a budget before it becomes a finance conversation.
Three actions this kind of breakdown enables:
Find the underused cheaper model. Gemini Flash handles summarization at roughly 40% lower cost than GPT-4o-mini at comparable quality in many workloads. A provider breakdown makes that migration candidate obvious.
Cap the experiment budget. Unguarded R&D spend is where runaway costs live. A misconfigured retry loop can burn hundreds in an afternoon — the dashboard flags it the same day instead of at invoice time.
Trace the spike. When a provider total jumps mid-month, project and operation labels let you trace it to a specific feature or worker in minutes — not after a week of digging through logs.
Model Meter is a TypeScript SDK. Connect it to your own traffic and the real numbers replace the example ones.
Model Meter on GitHub → · Join the waitlist for launch pricing →
Launch offer
