The cost stack of an AI product.

A single LLM call has three cost components.

Output tokens are 3–5× more expensive than input tokens because generation is inherently more compute-intensive than reading. Cached input is ~90% cheaper than fresh input.

The cost structure rewards short outputs, structured outputs, and repeated prompts. Your model's pricing is the boundary of your product economics.

The major models, side by side.

The cheapest model is often only ~6% the price of the flagship. Choose model by task difficulty, not brand — many production agents mix tiers.

One user task → 3 to 10 LLM calls.

An unconstrained agent task can cost $5–8. Budgets that quote per-token cost without accounting for fan-out are wrong by an order of magnitude.

One user task — Claude Sonnet 4.6 — full cost.

For 100,000 tasks/month: $28,200. Multiply by 12 months: $338K/year, just on inference. Caching can cut this by 70%.

The three caching tiers in production.

Stack them — they compose. Teams that implement all three see 70–90% production cost reduction relative to a naive implementation.

Hit rate by workload pattern.

An honest budget pessimistically assumes the lower bound. If your workload happens to be highly cacheable, you'll be pleasantly surprised; if it isn't, you won't have over-promised.

Median payback period by use case.

Source: 2026 cross-industry surveys (VentureBeat 1,100-engineer-and-CTO study; Digital Applied 100+ ROI data points).

For your group project: an AI feature with payback > 18 months is rarely funded. If yours lands there, find ways to halve the cost or double the value.

A 9-month payback example.

Hypothetical SaaS company, 8,000 monthly support tickets, $18.40 average resolution cost. Deploy an AI support agent.

When AI pays back fast, it's almost always because it displaces high-marginal-cost labour at high volume. When it pays back slowly, it's usually because a human still has to verify each output.

The course, in five lines.

1. Every software decision is an economic decision. Boehm's seven steps apply at any scale.
2. Cost lives in a fixed/variable × direct/indirect × non-recurring/recurring cell, across five lifecycle phases. 60–80% lives after launch.
3. Time value of money is the calculus of comparison. NPV is the verdict; IRR, payback, PI are the footnotes.
4. Sensitivity tells you which assumption to research; risk analysis tells you the probability of being wrong.
5. The AI era changes the coefficients, not the equations. Same Boehm framework, recalibrated for tokens, verification, and a re-allocated productivity curve.

If you walk away with only those five lines, you can defend any software decision in any room you'll enter for the next decade.

What's on the exam.

Closed book; one A4 cheat sheet (single-sided) permitted; non-programmable calculator.

Get your slides ready.

1. Finalise group-project presentation. Required slides: scope, FP/COCOMO estimate, cash flow + 2 alternatives, NPV/IRR/payback/PI, sensitivity, AI-cost dimension (bonus), recommendation.
2. Rehearse — 12 min talk + 5 min Q&A + 3 min peer evaluation.
3. Submit final project report PDF to /submissions/PROJECT/<team-name>/ before tomorrow's class.

What today bought you.

1. Token cost = input + output (3–5× input) + caching. Output tokens dominate.
2. Agents fan out 3–10×. Per-call cost ≠ per-task cost.
3. Caching is the highest-ROI optimisation. 40–90% in production is standard, not exceptional.
4. Payback depends on how much human labour you displace, and how cheaply.