Check out what I made: a custom cost report using Helicone API.

I've been using Helicone for a few months to track our OpenAI costs, and while the dashboard is great for a high-level view, I needed something more granular for our weekly FinOps sync. The built-in reports didn't quite slice the data the way I wanted—I needed to attribute costs per internal project and developer, across specific model families.

So, I built a script that pulls data via the Helicone API and generates a custom CSV report. It's been a game-changer for spotting waste, like catching a stray `gpt-4-32k` call in a project that should only be using `gpt-3.5-turbo`.

Here's the core of the Python script. It fetches request logs, aggregates them, and spits out a weekly summary.

```python
import requests
import pandas as pd
from datetime import datetime, timedelta

HELICONE_API_KEY = "your_key_here"
url = "https://api.helicone.ai/v1/request"

headers = {
"Authorization": f"Bearer {HELICONE_API_KEY}",
}

# Get last week's data
end_date = datetime.utcnow().date()
start_date = end_date - timedelta(days=7)

params = {
"from": start_date.isoformat(),
"to": end_date.isoformat(),
"limit": 1000 # Adjust as needed
}

response = requests.get(url, headers=headers, params=params)
data = response.json()

# Process and aggregate
df = pd.DataFrame(data)
# Add custom logic here to map 'userId' or 'customProperty' to internal projects
df['cost'] = df['totalCost'] # Use actual cost field from Helicone
df['model_family'] = df['model'].apply(lambda x: x.split('-')[0]) # Simple grouping

summary = df.groupby(['project_tag', 'model_family', 'userId']).agg({
'cost': 'sum',
'id': 'count' # Request count
}).rename(columns={'id': 'request_count'})

summary.to_csv('helicone_weekly_report.csv')
```

Key takeaways from running this:
* It exposed a pattern of low-utilization `gpt-4` calls during off-hours—turns out a scheduled job hadn't been downgraded.
* By tagging requests with a `project_tag` custom property, we finally got accurate cross-team chargebacks.
* We now combine this data with our AWS/Azure bills for a unified cloud cost view.

The API is straightforward. The main challenge was mapping Helicone's `userId` and custom properties to our internal directories. Once that was done, the aggregation was simple.

Has anyone else built custom reports? I'm particularly interested if you've found a way to correlate this LLM cost data with compute costs from your cloud provider for a total "cost per AI feature" metric.