Two context configurations are tested on identical tasks. A scores 0.91 correctness at 8,000 input tokens; B scores 0.90 at 3,000 tokens. The required correctness threshold is 0.88. What conclusion is best supported?

1. B is a promising lower-token option, subject to checking safety, latency, and important task slices.
2. Both configurations are equivalent because input tokens never affect cost.
3. A is always superior because more context is inherently better.
4. B must be rejected because every shorter prompt is less safe.