New Benchmark Shows Large Models Perform Poorly on Low-Latency Tasks, Highest Pass Rate Is 36%

Techub News reports that Realtime-Bench has been released to evaluate AI systems' performance in real-time interactive tasks, encompassing 50 web and game tasks. Tests show that OpenAI GPT-6 Astra achieved the highest pass rate in real-time conditions at 36%; when the clock is paused, high-inference models can approach 70% to 80%. This benchmark was constructed by the University of Texas at Austin, NVIDIA, Northwestern University, and other institutions, examining models' temporal state awareness, adaptive inference, and timing action capabilities. Models participating in the tests also include Anthropic, Google, DeepSeek, and others. (Import AI)