On 6 July 2026 a user asked Merlin a one-line question. The model thought for 54 seconds, spent all 2,048 of its 2,048 thinking tokens, and returned nothing.
The question, on the second turn of session 4afb8ec8:
"what is a structured intelligence assistant"
No files, no tools. Merlin sent the question down the direct-answer path, the fast lane for questions needing no evidence. The local model on the lane had thinking switched on. The model opened a thinking block, never closed the block, and the budget ran out inside. The answer section was empty. After 54 seconds the user got a typed error instead of an answer.
The question was easy. Stopping was the hard part.
A thinking budget sets how long a model reasons before answering. Nobody had decided what happens when the budget ends mid-thought, so the default decided for us, and the default was silence. Fifty-four seconds of thought on a one-line definition. I have sat in meetings like those. The difference is the meeting usually produces a sentence at the end.
I made two changes the same day. The direct-answer path now runs with thinking off. A definition question gains nothing from hidden reasoning, and the user pays for every reasoning token in waiting time. Reasoning stays on for planning and analysis, where the output gets checked against evidence. And when a model still returns thinking with no answer, Merlin salvages the most answer-like tail of the thinking text before giving up, and counts every salvage in a metric, so I see how often the net catches something.
The test: for every budget in your system, ask what the user sees when the budget runs out halfway through a thought. If the honest answer is "whatever the default does", nobody has decided.
Source: Spec 001, AC-FR005-06, written 6 July 2026. The run's logs have aged out of my log store. The specification entry written the same day records the numbers.
Top comments (0)