Time to first token, often abbreviated TTFT, is the wait before the answer starts. The model first processes the entire prompt, a phase called prefill, and only then begins generating. A long prompt, a large tool schema, or a cold cache all push this number up, even when the model generates quickly afterward.
TTFT is the number an interactive user feels. A model that generates two hundred tokens per second but takes eight seconds to start feels slow; one that starts in half a second and generates forty tokens per second feels responsive. Compare TTFT on your real context length, not on a short demo prompt.
For batch jobs, TTFT matters less than total throughput. For an agent loop, it compounds: every step pays a prefill, and a ten-step task pays ten starts. Measure it with the prompt you actually send. See throughput for the other half of the picture.