When you use a reasoning model like OpenAI’s o3 or DeepSeek R1, you see a clean, concise response — maybe 500 tokens. That’s the tip of the iceberg. It looks manageable. You multiply 500 tokens by the output price and think, “That’s cheap.”

Beneath the surface, the model generated thousands of invisible “thinking tokens” before producing your visible response. These tokens represent the model’s internal chain-of-thought reasoning — exploring solution paths, verifying logic, self-correcting. They’re billed at full output token rates but you never see them. A 500-token visible response can actually cost the same as a 10,000-token response.

Unlike standard models where output length is somewhat predictable, thinking token volume is entirely unpredictable. The same question asked slightly differently can trigger 500 or 15,000 thinking tokens. This makes cost forecasting for reasoning models extremely difficult. You can’t budget for what you can’t predict.

Imagine shipping costs worked like this: a 1 kg package costs $1, but a 2 kg package costs $4 (not $2). A 4 kg package costs $16. That’s quadratic scaling — doubling the weight quadruples the cost. This is exactly how transformer attention works. Every token must compare itself with every other token in the context, creating O(n²) complexity.

Quadratic scaling means the cost per token increases as context grows. The 128,001st token is more expensive to process than the 1st token, because it must attend to all 128,000 tokens before it. This is why providers charge surcharges for long contexts — the infrastructure cost genuinely increases non-linearly.

Both Anthropic and Google apply explicit price surcharges for long contexts. When your request exceeds 200K tokens, the per-token price doubles. This isn’t hidden — it’s documented on their pricing pages — but many developers miss it when estimating costs.

A RAG system that retrieves 50 documents averaging 5,000 tokens each loads 250K tokens per request. On Claude Opus, the first 200K costs $1.00 and the remaining 50K costs $0.50 (at the 2x rate). The surcharge adds 50% to the total input cost for that last 50K tokens. Keeping context under 200K becomes a concrete cost optimization target.

When you send the same prefix (system prompt, tool definitions, few-shot examples) across multiple requests, providers can cache the processed tokens and reuse them. Anthropic offers ~90% discount on cached token reads. OpenAI offers ~50% discount for prompts over 1,024 tokens. This is the single highest-ROI optimization — 10 minutes to implement, 45–90% savings on repeated prefixes.

For workloads that don’t need real-time responses, batch APIs process requests asynchronously at 50% off the standard price. Ideal for data processing, evaluation, content generation, and any task where a few hours of latency is acceptable.

Most cost estimates use the sticker price formula — input tokens times input price plus output tokens times output price. But the actual cost includes thinking tokens (which can be 10–20x the visible output), surcharges (2x above 200K), and misses discounts (caching, batching). Teams that budget on sticker price alone typically underestimate by 2–3x.

DeepSeek R1 handles 80% of reasoning tasks at negligible cost. o3 is the sweet spot for hard coding and math problems. o3-pro is for competition-level problems where accuracy is paramount. o1-pro is rarely justified outside research labs. The quality difference between tiers is real but narrow for most practical tasks.

Input tokens — your prompt, context, documents. Visible output tokens — the response you see. These are what the sticker price covers. For standard models (GPT-4o, Claude Sonnet), this is the full picture. For reasoning models, it’s just the tip.

Thinking tokens — 5–100x the visible output, billed at output rates. Quadratic scaling — each additional token costs more than the last. Context surcharges — 2x price above 200K tokens. Conversation accumulation — history grows with every turn. Tool schema overhead — 500+ tokens per tool, every request.

Chapter 4 takes everything from Chapters 1–3 and applies it to real-world bills. Five worked examples with March 2026 pricing: a consumer chat app, a customer support bot, a code assistant, a RAG system, and an autonomous agent fleet. The electricity bill analogy — moving from per-token thinking to monthly bill thinking.