A strong answer
Binary search uses O(log n) comparisons on sorted random-access data, while a linear scan can use O(n). For large arrays, the asymptotic difference usually matters. For small arrays, a scan may still win because its loop is simple, accesses neighboring memory, and can be easy for the processor to prefetch and vectorize. Binary search jumps among positions and has data-dependent branches.
The crossover depends on the language, element comparison, data size, cache state, and implementation. I would benchmark representative data with repeated runs, guard against compiler elimination, and report the target hardware and compiler settings. I would not claim a universal threshold.
I would also confirm that the input is sorted and define how duplicates should be handled. If I need the first equal value, I would use a lower-bound invariant rather than stop at any match.
Follow-up direction
Explain why the benchmark's input distribution and cache state resemble the actual workload, then compare branch and memory behavior if profiling tools are available.