AI Cost Lab

Simulators you can break and fix: cost per outcome, the token-efficiency levers, evals as the gate, agents going non-linear, and pricing what AI costs you. Three tracks, twenty steps, one fictive company (Northbeam) whose numbers add up. Built by Asbjørn Jørgensen; the seven levers it teaches are also the seven levels of Caveman.

Beginner: know what one request costs

1. What does one request cost?
2. Cost per outcome, not price per token
3. Where do the tokens actually go?
4. Tag it, or you can't fix it
5. Catch the leak before the invoice does
6. Your first lever: right-size the model
7. You can now answer "what does one request cost?"

Advanced: cut tokens without cutting quality

1. Cascading: cheap first, escalate on doubt
2. Prompt caching: the cheapest tokens are the ones you don't resend
3. Context trimming and output caps
4. Prompt compression, and how it fights your cache
5. Evals: the rule that makes every lever safe
6. Agents: step limits, budgets and circuit breakers
7. Batch and distillation: the deferred wins
8. The efficiency stack, in effort order

Expert: agents, trajectories and pricing

1. Trajectory cost is quadratic in steps
2. Compaction that respects the cache
3. Route per step, not per request
4. Evaluate trajectories on outcome and cost
5. Per-tenant budgets and fair use
6. Price what AI costs you
7. The credit cliff and the harness effect
8. From reporting the news to making it

Also: the theory map and the glossary. The lab needs JavaScript to run the simulators.