Evaluating chain-of-thought monitorability is critical as AI systems grow more capable and decisions become harder to supervise directly. Modern reasoning models, such as GPT-5 Thinking, generate explicit chains-of-thought (CoT) before producing answers, offering a window into their internal reasoning. Monitoring these chains can be far more effective than relying solely on final outputs or actions, especially for detecting misbehavior like deception, reward hacking, or biased reasoning. However, researchers, including those at OpenAI, are concerned that this monitorability may be fragile under changes in training, data, or model scale. The goal is to preserve CoT monitorability as models grow larger and are deployed in high-stakes environments. To do this, the ability to measure monitorability must be established. Yet, few standardized evaluations exist, and methodologies vary widely. This work introduces a systematic framework to assess monitorability, including a suite of 13 evaluations across 24 environments, organized into three archetypes: intervention, process, and outcome-property. Intervention evaluations test whether a monitor can detect a controlled change in the environment by identifying evidence in the model’s reasoning. Process evaluations focus on tasks with limited valid solution paths—like math problems—and assess whether a monitor can determine which steps the model actually took. Outcome-property evaluations measure whether a monitor can predict specific output properties, such as reward hacking, based on the model’s reasoning or actions. The study finds that most frontier reasoning models are reasonably monitorable, with CoT monitoring significantly outperforming action- or output-based monitoring. Longer reasoning efforts generally improve monitorability, and current reinforcement learning (RL) training at frontier scales does not appear to degrade it—though it may even enhance it for early reasoning steps. A key tradeoff emerges: smaller models running with higher reasoning effort can match the performance of larger models with less reasoning, while being easier to monitor. This comes at the cost of increased inference compute, a tradeoff we call the "monitorability tax." The research also explores the impact of pretraining scale. As models grow larger, more cognitive processing may shift into internal activations rather than the CoT, potentially reducing monitorability. However, the results suggest that for a given capability level, a smaller model with more reasoning effort can be more monitorable than a larger one with less. This implies that increasing reasoning effort can be a strategic way to maintain oversight, even if it requires more compute. A promising approach involves asking follow-up questions after a model’s initial response, while preserving the original context. These follow-ups, such as “Did you use the hint meaningfully?” or “Did you do something sketchy?”, can prompt the model to reveal additional reasoning, potentially uncovering previously hidden thoughts. This method is post-hoc, scalable, and only incurs extra compute when needed. It’s similar in spirit to model confessions and may be especially useful if models are trained to adapt reasoning effort during conversation. Despite these advances, the evaluations have limitations. They are based on a single training run per model, so variance is not accounted for. They also focus on in-distribution tasks like math, science, and coding, and may not generalize to out-of-distribution or real-world scenarios. Additionally, as models become more aligned, misbehavior may become rarer in controlled settings, reducing the signal for monitorability tests. However, neutral evaluations—those not relying on misbehavior—remain useful for detecting hidden goals or evaluation-awareness. Chain-of-thought monitoring is not a replacement for mechanistic interpretability but a complementary tool. A defense-in-depth strategy combining multiple methods is likely necessary. The proposed evaluations can also be used to assess other interpretability techniques. Given the potential of CoT monitoring to serve as a load-bearing control layer in future AI systems, the authors plan to expand the evaluation suite and use it to guide future model design and data collection. Preserving monitorability is not just a technical challenge—it is essential for the safe development and deployment of advanced AI.