Findings No. 11
A default reasoning budget can eat the entire output budget and return nothing
kimi-k2.6 with default reasoning spent the full output budget thinking in 62 of 64 cells and returned no JSON: $8.70 for zero scorable output. Measure what fraction of the budget your answers actually get.
A model with default reasoning enabled shares one output budget between its thinking and its answer, and nothing stops the thinking from spending all of it. In Stem Bench's chart-draft study, kimi-k2.6 with its default reasoning setting consumed the entire output budget thinking in 62 of 64 cells and returned no JSON at all: $8.70 of reasoning tokens for zero scorable output. The model never stopped deliberating long enough to answer.
The fix was to turn reasoning off for that model, matching the call shape the production app had already hard-coded for its own text-to-beat feature, which had learned the same lesson independently. The failed runs were archived as evidence rather than discarded, and the change was registered as a protocol amendment before any affected scored cell.
The general rule: when a provider defaults reasoning on, treat the output budget as contested territory, and measure what fraction of it your answers actually get.
Receipt: Stem Bench Study 1, REPORT.md Amendment 1 (commit 8c4201b); archived runs in results/archive/.