The bigger the question, the bigger the gain. On a one-line read, both ways cost about the same.
Our own bench, September 2026: Claude Sonnet and Opus on our test system, 66 databases. The conventional way is one tool call per operation on our own core, not a test of other vendors’ products. The times there: 90 to 120 seconds, 40 to 70 seconds, and 10 to 30 seconds asked again; they depend on the question, the model and the server.