A reasoning model can spend your whole budget thinking and return nothing.

What it looked like

The request succeeded, with no error, no timeout and no failed status. The page showed a single empty result card. To a visitor that looks like a broken product, and nothing in the logs says otherwise, which is what makes this failure easy to miss.

Where the tokens went

Reasoning models think in the open before answering, usually inside a think block, and that thinking is billed and counted against the same output allowance as the answer. Our generators cap output deliberately: 700 tokens for short hooks, 1,000 for ad copy, 1,200 for a script. Those caps are generous for the answers we want and far too small for a model that reasons first. The model hit the ceiling mid-thought and stopped before writing a single line of the actual output.

The model listing gave no warning. Ours had a name that suggested speed, and its price per token told us it was cheap. Neither said how many tokens it would spend before the answer began.

Why it showed as empty rather than as an error

Our parser strips think blocks before looking for JSON, which is correct. It then falls back to treating whatever remains as a single plain-text result when the JSON cannot be read, which is also reasonable. The two behaviours combined badly: strip the thinking, find nothing left, and return one item containing an empty string. The page rendered exactly what it was given.

• The strip only matched a closed think block, so a model cut off mid-thought left its reasoning in the output instead.
• The fallback returned an item whether or not any text survived the cleaning.
• An empty string is not an error, so nothing upstream complained.

Both are now fixed. Unterminated think blocks are stripped to the end of the response, and empty output raises a real error so the visitor sees a message they can act on instead of a blank panel.

What the models actually cost for this job

Our prompts are small. A 228-token system prompt plus a template and the user inputs comes to roughly 600 tokens in, against an output capped between 700 and 1,200. At that size the price differences between models are large in ratio and small in absolute terms, so price alone is a poor way to choose.

Measured at 600 input and 1,000 output tokens, September 2026. Reasoning models are left out: for this job, their real cost is a run that returns no answer.

The fifty-fold spread is real, but both numbers are small. Choosing the cheaper model is sensible once you have confirmed that it answers directly instead of reasoning first.

What to change

• Pick an instruct model for structured generation, not a reasoning model. Reasoning is valuable for hard problems and wasted on writing ten short lines to a fixed format.
• Do not read speed into a model name. Several models carrying words that suggest speed reason before answering.
• Treat empty output as an error, never as a result. A blank panel reads as a broken product.
• Strip unterminated think blocks, not just closed ones, because truncation is exactly when the closing tag goes missing.
• Check the token accounting after the first live run. Output tokens at or near your ceiling with nothing to show for them is the signature of this failure.

We build and run tools like these for other businesses as well: see our AI automation agency page.

Another error that hid its real cause, and how we traced it, is in our note on Instagram's "Container Publication failed".