The Hidden Costs of LLM APIs

Last quarter I helped a friend audit his AI writing tool. His pricing model assumed Claude Sonnet at $3 per million input tokens, $15 per million output. He'd projected $800/month at his current user count. His actual bill: $2,940. He wasn't doing anything unusual — no runaway loops, no compromised keys, no viral traffic spike. He was just budgeting from the pricing page instead of the invoice. This is the pattern. Every founder who builds on LLM APIs eventually discovers the sticker price is a floor, not a ceiling. The published rates from Anthropic, OpenAI, Google, and DeepSeek describe a laboratory clean transaction: one clean prompt, one clean response, no history, no retries, no tools, English. Nothing you ship looks like that. Your prompts carry system context, tool schemas, conversation history, RAG chunks, and a retry policy that sometimes double bills you when the provider hiccups. Your users type Arabic, or Turkish, or paste Python. Your staging environment runs 24/7. And once every six months, your model gets deprecated and you rewrite half your prompts to survive the migration. Here's what the invoice actually contains, and how to price it before it prices you. Tokenizer inflation is a tax on everyone except English speakers The pricing page quotes cost per million tokens. It does not tell you that a token is not a word, and that the ratio between them depends heavily on what language your users speak. For English, GPT and Claude tokenizers land near 1.3 tokens per word. For Arabic, it's closer to 3.5. For Chinese and Japanese, it can hit 2 tokens per character. Turkish, Thai, and Hindi all sit in the 2 3x range. This isn't a rounding error — it's a straight multiplier on your bill. A prompt that costs a US user $0.001 costs…