I've been using AI tools heavily since late 2024 — ChatGPT, Claude, Copilot, various API calls for side projects. Last month I decided to actually track what I was spending across all of them.
The results were... not what I expected.
The Setup
I exported billing from every AI service I use, tagged each charge by category (coding assistance, writing, image gen, API calls for apps), and threw it into a spreadsheet. Nothing fancy. Just honest accounting.
What I Found
Total spend: $287 in one month.
Here's the breakdown that shocked me:
- API calls from my own apps: $142 (49%)
- Coding assistants (Copilot, Claude Code): $68 (24%)
- ChatGPT Plus + occasional GPT-4 API: $43 (15%)
- Image generation: $21 (7%)
- Miscellaneous (embeddings, whisper, etc.): $13 (5%)
Almost half my AI budget was going to API calls from apps I'd built. Not tools I was actively using — background processes, chatbots, automations. Things that were burning tokens while I wasn't even looking.
The Hidden Cost: Tokens You Don't See
This is the part nobody talks about. When you build with LLM APIs, the meter is always running. A chatbot with a long system prompt? That's eating input tokens on every single message. A coding agent that reads your whole codebase into context? Thousands of tokens before it writes a single line.
I was running a small automation that summarized emails. Seemed harmless. It was costing me $3/day because each email thread was getting stuffed into a 16k context window.
What I Changed
1. I started actually monitoring token usage in real-time.
I put TokenBar in my menu bar so I could see token counts as I worked. Sounds simple, but it completely changed my behavior. When you can see 4,000 tokens disappearing on a single prompt, you start writing tighter prompts. It's like watching your electricity meter while running the AC — suddenly you care.
2. I set budget alerts on every API key.
OpenAI and Anthropic both let you set spending limits. I should have done this from day one. Now I get pinged when any project crosses $20/month.
3. I killed the background jobs that weren't earning their keep.
That email summarizer? Replaced it with a simple rule-based filter that handles 80% of cases, with AI only kicking in for the ambiguous ones. Went from $90/month to $11.
4. I started caching aggressively.
Same prompt + same input = same output. There's no reason to hit the API twice. I added a simple Redis cache in front of my most common calls and cut redundant requests by ~60%.
The Uncomfortable Truth
Most developers I talk to have no idea what they're spending on AI. They know their subscription costs ($20 for ChatGPT, $10 for Copilot), but the API spend is a black box. And it adds up fast once you're building things.
We've gotten really good at monitoring server costs, database queries, and CDN bandwidth. But token usage is still the Wild West for most teams.
My Spend Now
After a month of actually paying attention: $147. Almost exactly half of what I was spending before, with zero loss in functionality. The savings came entirely from eliminating waste — not from using AI less.
The Takeaway
Track your tokens like you track your AWS bill. The tools exist. The awareness doesn't — yet.
If you're building with LLMs, the biggest cost optimization isn't a better model or a cheaper provider. It's knowing where your tokens are actually going.
What's your AI spend looking like? Anyone else been surprised when they actually added it up? Would love to hear what others are seeing.
Top comments (0)