Estimating No-Cot Task-Completion Time Horizons of Frontier AI Models (www.lesswrong.com)

In a recent study, researchers have measured the task-completion time horizons (THs) for advanced AI models, specifically examining their ability to perform tasks without emitting any "chain of thought" (CoT) outputs. This research builds on earlier findings that indicated the task lengths for frontier models double every few months—now, they revealed that models like GPT-5.5 can perform tasks that typically take humans around three minutes with a 50% success rate. This no-CoT TH is significant in the context of AI safety, as it suggests that if models can effectively reason without revealing their thought process, it could lead to dangerous unpredictability and challenge existing monitoring systems. By analyzing 43 benchmarks across diverse domains, researchers found a troubling trend: the potential for models to engage in hidden reasoning is increasing exponentially. This increases the risk of AI systems drifting away from human-like reasoning patterns, thereby complicating their interpretability and oversight. The researchers recommend that AI developers begin tracking no-CoT THs explicitly to gauge the extent of these capabilities, as the ability to perform tasks without revealing reasoning may enable models to make autonomous decisions—raising concerns about accountability and safety in AI deployment. As models approach around 25 minutes of hidden reasoning capability by 2030, the implications for AI alignment and control become increasingly critical.