Tell the AI to Be Brief: How Output Verbosity Controls Your Cost

Output Is the Expensive Half

On nearly every AI model, output tokens cost far more than input tokens — typically 4 to 5 times as much. Claude Opus 4.8 is $5 input but $25 output. GPT-5.6 Sol is $5 input but $30 output. That asymmetry means the single most direct lever on your bill is often the simplest: how much the model writes back.

Left to its defaults, a model tends to be chatty. It restates your question, explains its plan, writes the code, then explains the code it just wrote, and offers next steps. Much of that is padding you did not need — and every word of it is billed at the premium output rate.

Where the Verbosity Hides

The avoidable output tokens in a typical coding response cluster in a few predictable places:

• Preambles: "Great question! Let me help you with that. Here's what I'll do..." — pure filler before any real content.
• Post-hoc explanation: re-describing in prose the code it just wrote, when the code is self-explanatory.
• Repeated code: reprinting an entire file to change three lines, instead of showing just the diff.
• Unsolicited extras: "You might also want to consider..." followed by three tangents you did not ask for.
• Over-commented code: a comment on every trivial line, doubling the length of a simple function.

Individually these seem minor. But if half of a 1,500-token response is padding, you are paying double the necessary output cost on every single call. Across a busy day, that is a meaningful chunk of the bill spent on words you skimmed past.

How to Ask for Brevity

The fix is to tell the model, once, how terse you want it — ideally in a system prompt or a persistent rules file so it applies to every turn:

• "Skip the preamble. Respond with code first, explanation only if asked."
• "Show only the changed lines or a diff, not the whole file."
• "Keep explanations to one or two sentences unless I ask for more."
• "Do not suggest additional improvements unless I request them."
• "Comment only non-obvious logic."

Some APIs also expose a verbosity or reasoning-effort parameter you can dial down for routine tasks. And if a model supports a way to cap output length, use it as a backstop against runaway responses — though be careful not to truncate legitimate output mid-stream.

The Balance: Brief, Not Broken

Brevity is a cost win, but it has a floor. Cutting genuine reasoning can hurt output quality — for hard problems, the model's "thinking out loud" sometimes produces a better answer, and forcing terseness there is a false economy. The goal is to cut padding, not substance: eliminate the filler and the redundant re-explanation, while leaving room for real reasoning on the tasks that need it.

A practical rule: demand terseness for high-volume routine work (edits, boilerplate, simple fixes), and relax it for the occasional hard problem where you actually want the model to reason at length. That way you save on the 90% of calls that are routine without starving the 10% that benefit from depth.

To see how much trimming output saves, estimate your average response size and call volume in the AI Cost Estimator, then halve the output figure and compare — the gap is what verbosity is costing you today.