Cekura Bench gives founders and operators a public way to compare voice AI before they ship a phone agent. It runs speech-to-speech models, speech-to-text, text-to-speech voices, and deployed voice agent platforms on the same scenarios, then reports reliability, reply time, cost, and voice quality. Results are at benchmarks.cekura.ai. Cekura, the YC Fall 2024 company behind the boards, already sells simulation, evaluation, and production monitoring for voice and chat agents. This release covers more than 50 models from OpenAI, Google, Nvidia, xAI, Inworld, Gradium, AssemblyAI, Fish Audio, and others. The calls include accents, interruptions, changes of mind, and background noise, and a separate track uses recordings of real people doing tasks. Reliability, speed, cost, turn-taking, and voice quality are led by different systems. The bench extends work Cekura already does for customers: simulate calls before launch, score them, and monitor production traffic. CEO Tarush Agarwal announced it. He co-founded the company with Shashij Gupta and Sidhant Kabra, who met as undergraduates at IIT Bombay.