Cost Series 05 — Output Discipline

Working notes for a planned episode — an outline, not an article. Details will shift when it gets scripted. Series index: The Bill.

The axis nobody counts: output is ~5× input and can't be cached. Every wasted output token is the most expensive kind.

The hook

You can cache the entire input and route to a cheap model — and still get a big bill, because the model is writing a novel when you needed a sentence. Output is the one axis caching can't touch, and it's the priciest per token. Discipline here is pure margin.

What the viewer walks away with

• Output ≈ 5× input and uncacheable — the least-optimized, most-expensive axis
• Ask for less: structured/bounded outputs, stop sequences, "answer only" instructions, no regurgitating the input
• Reasoning/thinking tokens are output too — budget them; more thinking isn't free
• Shorter outputs also cut latency — the rare optimization that helps cost and speed at once

Beats

• The cached-input, cheap-model, still-expensive request — reveal it's all output
• The levers: schema/format constraints, max tokens, stop sequences, don't echo the prompt
• Thinking-token budgets (callback to the fleet's concurrent-workers gate lessons) — output you can't see is still output you pay for
• Close: caching is the input story, discipline is the output story — you need both

Demo ideas

• Same answer, verbose vs bounded output — token and cost delta
• A thinking-budget sweep: quality vs output-cost curve, where the knee is

Source material

• The series index — the output multiplier
• The fleet's concurrent-workers gate lessons — thinking-token budgets that already bit us
• The token-usage analysis — output-share of real bills

Open questions

• How hard to push "constrain the output" before it hurts quality — where's the line
• Treat reasoning tokens as their own mini-topic or fold into output