AI systems still fail most tasks where speed matters, new test finds

AI systems are still poor at jobs where seconds count. A new test called Realtime-Bench found that the best model passed just 36% of timed tasks under real-world conditions. Those tasks include buying tickets, shopping online and playing games. The same models did far better when the clock was paused, which suggests the problem is speed, not skill.

What the test found

The top score on the real-world setting, called the wall clock, went to GPT-6 Astra (high), at 36%. That means even the leader failed most tasks. With the clock paused, models set to think harder scored closer to 70%, and some reached 80%. In plain terms, the models can often do the work. They are just too slow when the world keeps moving.

How the test works

Realtime-Bench has 50 tasks. Thirty are web tasks in four groups: travel, e-commerce, service and security. Twenty are games: arcade, puzzle, strategy and simulation. The researchers say the tasks are ones where a delayed action directly hurts the result.

It checks three skills. The first is tracking changes in the world. The second is thinking for less time when time is short. The third is picking the right moment to act. A ticket buyer, for example, must know whether tickets are still available, stop deliberating as the window closes, and decide whether to wait or buy now.

• Wall clock: real-world delays, as a user would experience them.
• Paused clock: time stops, to measure basic ability.
• Token clock: time moves forward as the model produces more text, so longer thinking means later action.

Who was tested

The authors tried models from OpenAI, Anthropic and Google, plus Jev 1.13, Deepseek (V4.1) and Moonshot AI (Kimi-K3). The work came from researchers at the University of Texas at Austin, Ohio State University, the University of Illinois at Urbana-Champaign, NVIDIA, Northwestern University and Steadyworks.

What to do

Be careful about letting an AI agent handle time-sensitive jobs for you, such as ticket drops or online auctions. Slow thinking can cost you the purchase. For tasks with no clock pressure, the paused-clock results suggest these tools are much more capable.

What happens next

The authors say the results show it will be a while before AI systems can play games or run online auctions for you. A public leaderboard and the research paper are available. The material does not say how quickly models might improve or how any specific consumer product would score.