Most startups integrating LLMs run into the exact same wall around month 6.
User growth looks great. ARR is going up. But your OpenAI/Anthropic bill is growing 3x faster than your MRR. Suddenly your gross margins are negative, and you have no idea why.
I've talked to dozens of founders this year. Almost everyone starts the same way: one global API key, no caching, and a "we'll figure out costs later" mentality. Later is usually when Stripe fails to cover the API bill.
The problem isn't the model pricing. GPT-4o mini and Claude 3.5 Haiku are cheap. The problem is lack of visibility.
When a customer complains about an issue, your support team runs an agent loop. When a user uploads a PDF, your RAG pipeline chunks and embeds 50 pages. Who paid for that? Which customer is actually profitable?
Usually, 20% of your power users are burning 80% of your API budget, while the rest are subsidizing them. But without per-tenant cost attribution, you can't tell them apart. You can't adjust your pricing tiers.
If you don't track costs per user, you are flying blind.
Start tracking per-tenant usage early. Even logging token counts to your database is better than nothing.
fwiw, if you don't want to build it yourself, I built LLMeter (https://llmeter.org?utm_source=devto&utm_medium=article&utm_campaign=2026-04-15-devto-ai-costs-revenue). It's an open-source dashboard that tracks LLM API costs by model, by user, and by day. Handles OpenAI, Anthropic, DeepSeek, and OpenRouter.
Stop guessing where your margin went.
Top comments (0)