Terms  /  Reasoning Cost  /  Context Compensation
02 · Reasoning Cost

Context Compensation

The model using additional reasoning to make sense of poor, bloated, or unstructured context.

Context Compensation is the mechanism behind Reasoning Tax. When a model receives unclear or contradictory input, it doesn't give up — it compensates. It generates internal reasoning to resolve ambiguities, hypothesize about intent, and filter signal from noise. This compensation is a testament to model capability, but it's also expensive and unreliable. The model is essentially guessing what you meant, and those guesses cost reasoning tokens. When you see a model produce a long chain of thought before a simple answer, you're often watching Context Compensation in action.

Example
A user sends a prompt that says: "Make it better." The model has no clear criteria for "better," so it compensates: it reasons about what "better" might mean in this context, considers multiple interpretations, selects the most likely one, and then executes. This compensation might produce 800 tokens of internal reasoning for a task that, with a clear prompt ("reduce the function's time complexity from O(n²) to O(n log n)"), would have required 200. The extra 600 tokens are the cost of Context Compensation.

More in Reasoning Cost