Researchers find AI models struggle with real-time interactive tasks
Reported by Import AI1 source ↓
A new benchmark finds that frontier AI models perform much worse when tasks require actions under real-world latency than when the environment can be paused. The newsletter also summarizes a Defense Intelligence Agency report on threats to satellites.
Key points
- GPT-6 Astra (high) had the top wall-clock pass rate, at 36%
- With a paused clock, high-reasoning models approached 70% to 80% success
- Realtime-Bench contains 50 interactive game and web tasks
- The DIA report describes China and Russia developing capabilities to disrupt or damage satellites
Details
Realtime-Bench was developed by researchers from the University of Texas at Austin, Ohio State University, the University of Illinois at Urbana-Champaign, NVIDIA, Northwestern University and Steadyworks. It evaluates whether models can track changing conditions, adjust deliberation to available time and time their actions. Its 50 tasks include 30 web tasks across travel, e-commerce, service and security, and 20 games spanning arcade, puzzle, strategy and simulation.
In the newsletter’s account of the benchmark results, GPT-6 Astra (high) had the best pass rate under the wall-clock condition, at 36%. When the clock was paused, high-reasoning models approached 70% to 80% success. The benchmark also has a token-clock condition, where the environment advances as a model generates tokens, so longer responses delay action.
The newsletter also summarizes the Defense Intelligence Agency report Challenges to Security in Space. The report describes China and Russia as developing capabilities to disrupt or destroy other countries’ space services, including jamming, cyber operations, directed energy and systems that could threaten satellites.
Written from coverage as of
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh