A model's measured ability to produce correct answers on tasks requiring several dependent steps, such as arithmetic word problems, code tracing, or logic puzzles.

The most load-bearing phrase in AI marketing, used without a benchmark, a task, or a number attached, so that it can mean whatever the reader is hoping it means.

• Unqualified Claim: "Reasoning capabilities" with no task named and no score reported describes nothing measurable.
• Borrowed Credibility: A word carrying centuries of philosophical weight, applied to next-token prediction.
• House-Cleaning Reality Check: Like following a recipe correctly once and telling guests you have developed "culinary reasoning capabilities."

Short Circuit (1980s) where Johnny 5 flips through encyclopedias at high speed shouting for more input, and everyone in the room decides this proves he is thinking.

How well a model handles problems that take several steps to solve.

“The new model demonstrates substantially improved reasoning capabilities across enterprise workloads.”

“The model scored 71 percent on our set of 200 multi-step tax questions, up from 58 percent.”