Ivan Lopez

- Tokenomics in plain terms: why output costs 5x input, and why you pay to re-read the whole conversation on every step. One silly one-line question cost us $1.33, because ~63,000 tokens of context were read before our 8.

- The cache: a real 4-hour, 204-step coding session cost about $22 with caching and would have cost ~$320 without it. 70% of the bill was the model re-reading what it had already seen.

- Why cheaper tokens don't mean a cheaper bill: newer models can spend their whole discount thinking longer at max effort, and a model at half the per-token price can cost 1.5x more per finished task.

- Our biggest lever: agent orchestration. One planner on a premium model, specialists on the cheapest model that can do each job, and hard rules learned the expensive way (like the day 12 sub-agents burned our entire usage limit in 10 minutes with no results).

You'll leave with a checklist for Monday: measure cost per finished task, not price per token; manage context; and pick the model and effort level per job.