It's Wednesday, 6:14am. CFO Priya's email subject line is in all caps: 'WHAT HAPPENED TO INFERENCE COST'. Mateo forwards it to you at 6:17 with the words 'kai read a twitter thread last night'. You scroll the dashboard. Yesterday's spend is 3.1x Monday's. Same traffic, same prompts, same models — different dial. Kai, full of conviction after some thread about 'always max out reasoning_effort for production', set every single Sentinel prompt to `reasoning_effort=high` at 11pm. The real-time dashboard prompts now take 12 seconds. Doctors are tabbing back to Notion because the summary still isn't there when they leave the room. The nightly batch summarizer, which has all night to think, is also at `high` — fine, but pointless overkill. Today's lesson is the dial itself: reasoning effort, extended thinking, chain-of-thought. They are dials, not defaults. Every production prompt has a binding constraint, and the dial follows from the constraint — not from the thread.
You'll walk out able to
- Pick the reasoning_effort level on purpose, against a named binding constraint
- Distinguish tool-use latency from thinking-budget latency in a session trace
- Argue for opposite dials on opposite jobs (real-time vs nightly batch) using the same triangle
- Recognize when `none` (no reasoning) is the correct production pick
- Compare explicit chain-of-thought against silent extended thinking on cost-quality grounds
- Write eval cases that can actually distinguish between two adjacent dial positions
Your first task
Kai's 'everything is high' batch.
You sit down next to Kai at 8:30am with two coffees. 'Walk me through what you ran last night.' He opens the terminal. He's been running the same three dashboard prompts in a loop to 'feel the quality'. The latency log scrolls. He's confused — he expected high reasoning to make the model smarter, not slower. He points at one line and says 'see, that 14 seconds is the tool call, not the thinking — it's the database lookup'. Mateo, watching from the doorway, raises an eyebrow.
What you'll answer
Read Kai's session below. What's the diagnosis you give him?
Free · no card · we save your progress
Save your progress & earn the certificate
You've seen the opening — keep going free. 7 tasks · auto-graded · finish in 60 min and the certificate is yours. Free forever · no card.
No password. To sign in on your phone, request an emailed code on the sign-in page.
Already started? Sign in with an emailed code instead.