$113,421 in a single month

This is what production AI costs when nobody's watching.

A 4-person team posted their Anthropic invoice.

Agentic systems don't make just one API call per task.
They read context, plan steps, call tools, hit errors, retry.

Each step is a separate call to Opus at $25 per million output tokens.

One user instruction can trigger 20+ calls before it's done.

A lot of engineers have no idea what a single task costs end-to-end.

- They don't know which prompts trigger the longest loops
- They don't know how many silent retries are happening in the background
- They can't tell which tasks could run on a smaller model without losing quality

Frontier models are genuinely impressive.

But agentic systems don't make one call.. they make dozens.
Every single day.

And most teams aren't watching the meter.

If you're running agentic workloads in production, start tracking what individual tasks actually cost before your next invoice does it for you.