Realtime-Bench: GPT-6 Astra tops wall-clock test at 36%

Realtime-Bench arrives as businesses increasingly look to AI agents for tasks that involve deadlines and changing conditions. The benchmark's design separates three skills that real-time work demands: tracking changes in the world, adjusting how long a model deliberates, and timing actions correctly. The gap between wall-clock and paused-clock results suggests that current models can often solve these tasks when given unlimited time, but struggle when actions must be committed under real latency. That distinction matters for anyone planning to deploy agents in customer-facing or transactional settings, where a missed window can mean a failed purchase or a lost auction. The researchers frame the benchmark as a proxy for where agents will break, pointing to specific task types where caution is warranted.