Cekura Bench

Cekura Bench is an AI developer tool that benchmarks real-time speech-to-speech models by running them through actual phone calls, so teams can choose a voice model on evidence rather than vendor claims. Each model is tested on 82 real call scenarios, 59 appointment bookings and 23 Medicare intake calls, and every scenario is run three times to expose inconsistency. The results are scored across reliability (a pass-cubed metric that rewards models passing all three runs), success rate, data accuracy, stalled calls, response time, and cost per minute. What makes the benchmark usable rather than just a leaderboard is transparency: every call is published with its transcript, tool calls, and detailed scores, and you can run any two models head to head. A failure-analysis view shows exactly which scenarios break a given model and how different models fail differently, which is often more decisive than an aggregate score. For engineers and product managers asking what is the best AI for a production voice agent, Cekura Bench turns that question into comparable, inspectable data. The benchmark itself is free to browse; the per-minute figures shown are each model provider's own pricing, not Cekura's.

Comparing and selecting real-time speech-to-speech AI models for voice agents using transparent, call-based benchmark data.