In the semiconductor sector, a concept known as Moore’s Law, speculated since 1975, predicted that processing power would double approximately every two years, accompanied by declining costs. This progression transformed massive, room-sized computers into small yet powerful devices like the smartphones we carry today. Interestingly, the law itself motivated continuous innovation, effectively becoming a self-fulfilling principle. Now, an analogous phenomenon appears to be emerging specifically in artificial intelligence development. According to recent findings by researchers at METR, the capability of AI systems to perform prolonged tasks is currently doubling roughly every seven months, although the exact timing depends somewhat on the nature of the tasks themselves. This indicates a strikingly rapid increase in AI’s capacity to engage in longer and increasingly complex assignments.
To illustrate this progression, the researchers conducted comprehensive tests to evaluate how long various AI models could persistently perform complex tasks before encountering failure. The criterion they used to assess AI performance was the length of time a skilled human worker traditionally requires for the exact same task. Their examination included influential models spanning several years, from GPT-2, introduced in 2019, to Claude 3.7 Sonnet, expected in 2025. The selected testing processes involved demanding tasks that typically occur in fields such as software development, cybersecurity operations, and complex logical reasoning exercises. Through these detailed assessments, researchers uncovered impressive results: the best current AI systems have now attained proficiency enough to successfully handle work estimated to require an experienced human about one hour to complete.
A useful way to envision this research is to consider it akin to tracking an athlete’s measurable progression over time. In such a scenario, one might typically measure two basic parameters: how high athletes can leap and how long they can sustain continuous running before succumbing to fatigue. Traditional AI performance benchmarks have usually emphasized evaluating how well systems perform short-duration, isolated tasks, akin to assessing skill at high jumps. Alternatively, METR’s new research uniquely measures the endurance aspect of AI performance—tracking precisely how long AI systems can remain effectively operational and focused without failing, much like measuring how long an athlete runs efficiently before exhaustion.
METR refers to these observations as a type of “Moore’s Law for AI agents,” highlighting an exponential pattern. In practical terms, approximately every seven months, artificial intelligence systems double the amount of time they can effectively sustain task performance. If current trends persist, we might witness AI models completing daily eight-hour job tasks routinely by 2026. Shortly afterward, around 2027, AI could feasibly handle work that ordinarily spans two to three days. By 2028, AI systems might realistically manage tasks equating to an entire traditional 40-hour workweek without interruption or significant decline in performance quality. Expanding even further, projections forecast AI capabilities handling month-length assignments around 2029. Stunningly, by approximately 2031, artificial intelligence might possess enough endurance and proficiency to complete continuous projects equivalent in scope to a regular 50-week professional working year.
More accelerated progress emerges when examining software engineering tasks. Using the SWE-bench Verified data set, which represents actual programming and software development challenges, researchers reported an even more aggressive growth trajectory, demonstrating task endurance doubling in less than three months. Furthermore, data exclusively from models developed between 2024 and 2025 show a pronounced acceleration in AI capabilities, possibly shifting the previously mentioned projected milestones forward dramatically—by nearly two-and-a-half years.
Not all experts share uniform confidence in or acceptance of this exponential forecast. Researcher Tamay Besiroglu offered criticism after conducting a similar analysis regarding AI performance in chess, which resulted in significantly elongated projected timelines spanning decades for major milestones. Both Besiroglu and METR’s Megan Kinniment point out that while this exponential increase may be credible within certain well-defined domains, it is inappropriate to universally extrapolate that AI will soon manage all hour-long tasks with equal success. Rather, these dynamics will likely apply depending on the specific conditions and domains where AI performance is evaluated.
Additional experts raised concerns about the assumptions behind these findings. Some skeptical voices expressed doubts about the practical implications of merely achieving a 50% success rate on given tasks, questioning whether such a threshold realistically translates into genuinely reliable or usable results in professional settings. Similarly, others suggested there might be diminishing returns on future advancements, predicting scenarios where substantially greater computational resources may be required to achieve increasingly marginal improvements in capability—an effect mathematically predictable in many technological fields.
Even acknowledging the complexities and potential limitations, the profound implications of these research outcomes remain difficult to overlook. Particularly for highly educated and skilled white-collar professionals, this new data signals an increasingly tangible yet complex path ahead. AI systems may not imminently replace entire job roles overnight, but their accelerated improvement in executing extended tasks typical of professional duties is evident. In short, mechanized assistance may not immediately assume total job responsibilities, yet rapid progress in AI abilities continues, particularly in areas essential for white-collar occupations. Continued vigilance and careful adaptation by workers and organizations alike are strongly advisable in response to these fast-evolving technological developments.