Explicit reasoning can assist language models in generating code, yet it introduces a practical risk: the model may consume most of its output budget on reasoning and never produce usable code. We term this phenomenon reasoning‑trace overflow and study it across seven open‑weight models and six code‑generation benchmarks, all constrained by a fixed 32k‑token output limit. We measure the proportion of tokens spent on reasoning, its impact on pass@1, and whether lightweight inference‑time controls can alleviate overflow.\n\nThe results show that the value of reasoning is highly conditional. On simple function‑level tasks overflow is rare, but on harder benchmarks and for certain model families it becomes common—on several model‑benchmark pairs more than half of the generations exhaust the full 32k‑token budget while still reasoning. Most complete overflows are linked to repetitive reasoning loops, though substantial long reasoning also occurs without looping.\n\nConsequently, disabling reasoning can be competitive for overflow‑prone models because it guarantees an answer, while stronger models still benefit from reasoning on difficult algorithmic and code‑understanding tasks. We compare two control strategies: prompt steering can markedly reduce overflow for some models but is inconsistent, whereas budgeted transitions offer more direct control, improving both performance and token efficiency for overflow‑prone models.\n\nThese findings suggest that explicit reasoning should be treated as a controllable resource in code generation rather than an always‑on capability.\n\nReview: The paper provides a thorough empirical analysis of the trade‑off between reasoning depth and token budget, offering practical guidance for deploying reasoning‑enhanced code generators.