Each model is probed with a synthetic streaming request. Latency is total completion time; TTFT is time to the first streamed token. History is kept from live probes — no synthetic data.
Loading…