Nobody thinks about their electricity bill in kilowatt-hours. You think: “My bill was $180 this month.” AI costs work the same way. Per-token pricing is useful for understanding the unit economics, but what matters is the monthly bill at production volume. This chapter translates everything from Chapters 1–3 into real monthly costs for five common workloads.

96% of enterprises report AI costs exceeding initial projections. Teams routinely underestimate by 3x or more when planning on sticker price alone. The gap comes from hidden multipliers (Chapter 3), conversation history accumulation, retry loops, and the difference between prototype volume and production volume.

A customer support chatbot handling 10,000 conversations/month. Average conversation: 5 turns, 500 input tokens + 300 output tokens per turn. System prompt: 1,500 tokens (sent every turn). RAG context: 2,000 tokens per turn. Total per conversation: ~22,500 input + 1,500 output tokens.

Same workload, 21x cost difference. GPT-4o mini at $43/month vs Claude Sonnet at $900/month. For most customer support tasks — FAQ answers, order status, policy lookups — the budget model performs comparably. The expensive model only justifies its cost for complex troubleshooting or nuanced conversations.

20 developers, each making ~50 code requests/day (completions, explanations, refactoring, debugging). Average request: 2,000 input tokens (code context + prompt) + 1,000 output tokens (generated code). System prompt: 800 tokens. Total: 1,000 requests/day × 2,800 input + 1,000 output.

At $408/month for 20 developers, that’s $20/developer/month on GPT-4.1. If each developer saves just 30 minutes/day (conservative for a good code assistant), that’s 10 hours/month of saved time. At $75/hour loaded cost, that’s $750 in saved productivity per developer — a 37x return on the $20 API cost.

A RAG system answering questions over a corporate knowledge base. 1,000 queries/day. Each query retrieves 5 document chunks averaging 1,000 tokens each (5,000 tokens of context). System prompt: 1,000 tokens. User query: 100 tokens. Output: 500 tokens. Total per request: 6,100 input + 500 output.

RAG is input-heavy (92% input tokens), which makes it relatively cheaper per request. But the cost scales with retrieval depth — retrieving 10 chunks instead of 5 doubles input cost. The key optimization is retrieval quality over quantity: better chunking and reranking means fewer, more relevant chunks, which reduces both cost and improves answer quality.

7 AI agents handling software engineering, data analysis, and operations tasks around the clock. Each agent averages 50 tasks/day. Each task requires 5–10 LLM calls (planning, tool use, execution, verification). Average per task: 15,000 input + 8,000 output tokens. Context grows with each call in a task.

These estimates assume well-constrained agents. In practice, agents encounter errors, retry, and sometimes enter doom loops. A real Series B fintech startup reported their agent fleet cost $4,200/month in LLM fees + $2,800 in infrastructure. Without monitoring, teams lose $8,000–23,000/month to undetected waste.

$5–2,000/month depending on scale. Includes hosting, databases, vector stores, monitoring tools, and compute for non-LLM processing. For a simple chatbot, this might be $5–50/month on a managed platform. For an agent fleet, $500–2,000/month for orchestration infrastructure.

$5–50,000+/month depending on volume and model choice. This is the variable cost that scales with usage — everything we’ve calculated in the examples above. For light usage, $5–50/month. For heavy production, $1,000–50,000+/month.

10–30% annually on top of direct costs. Includes engineering time for prompt optimization, monitoring setup, incident response, model migration when providers change APIs, and ongoing evaluation. This is the cost most teams forget — the human labor required to keep AI systems running well.

Monthly totals tell you what you spent. Cost-per-task tells you whether it was worth it. The formula: (total tokens × price) / successful completions. If your agent completes 80% of tasks successfully, the cost per successful task is 25% higher than the raw per-task cost. Failed tasks still consume tokens.

Every AI task must generate more value than it costs. A customer support resolution that saves a $15/hour support agent 10 minutes is worth $2.50. If the AI resolution costs $0.15 (budget model), that’s a 16x return. If it costs $5.00 (premium model), you’re losing $2.50 per resolution. The math changes everything.

1. Calculate tokens per task (input + output + thinking). 2. Multiply by daily volume × 30. 3. Apply the full cost formula from Chapter 3 (including thinking tokens and surcharges). 4. Add infrastructure costs (40–60% of LLM cost). 5. Multiply total by 2–3x for safety margin. 6. Calculate cost-per-successful-task and compare to value generated.

We’ve covered the complete token economics story: what tokens are (Ch 1), how they’re priced (Ch 2), the hidden multipliers (Ch 3), and what real bills look like (Ch 4). Chapter 5 shifts to the GPU and infrastructure layer — the real estate analogy of renting vs buying vs building.

Customer support: $43–900/month depending on model. Code assistant: $20/dev/month with 37x ROI. RAG: $36–774/month. Agent fleet: $3,800–5,700/month. The three-layer cost stack (infra + API + overhead) means LLM cost is only 40–60% of total. Cost-per-successful-task is the metric that matters. Always multiply estimates by 2–3x.