Realtime-Bench finds top large model pass rate at 36% in low-latency tasks

Realtime-Bench has been released to measure how AI systems perform in real-time interactive settings, using 50 web and game tasks. The benchmark found that OpenAI GPT-6 Astra posted the highest pass rate under real-time constraints at 36%. When the clock was paused, high-reasoning models could reach roughly 70% to 80%. The benchmark was built by researchers and institutions including the University of Texas at Austin, NVIDIA, and Northwestern University. It evaluates time-state awareness, adaptive reasoning, and the ability to act at the right moment. Models from Anthropic, Google, and DeepSeek were also included in the testing, according to Import AI.