Measuring AI Ability to Complete Long Software Tasks
Authors: Thomas Kwa, Ben West, Joel Becker, Amy Deng, Katharyn Garcia, Max Hasin, Sami Jawhar, Megan Kinniment, Nate Rush, Sydney Von Arx, Ryan Bloom, Thomas Broadley, Haoxing Du, Brian Goodrich, Nikola Jurkovic, Luke Harold Miles, Seraphina Nix, Tao Lin, Chris Painter, Neev Parikh, David Rein, Lucas Jun Koba Sato, Hjalmar Wijk, Daniel M. Ziegler, Elizabeth Barnes, Lawrence Chan
Published March 18, 2025
Read the paper: arXiv 2503.14499 · PDF
Why it matters
Benchmark scores are hard to translate into what a model can do on the job. The time horizon turns them into a unit people understand: the length of a task, measured by how long it takes a human expert. The authors timed people with relevant expertise on a mix of RE-Bench, HCAST and 66 new shorter tasks, then found the task length at which each model succeeds 50% of the time.
Their explanation for the growth is useful too: models got more reliable, better at recovering from mistakes, and better at logical reasoning and tool use. The authors are clear about the limits. The tasks are software and research tasks, how well they represent real work is an open question, and the five-year forecast in the abstract is an extrapolation that holds only if the results generalize.
How to use it
Use the time horizon to size the work you give a coding agent. Tasks that take a skilled person minutes are a reasonable place to hand off whole; longer ones are better split into smaller steps, with checks between them. Read the 50% in the name carefully: at the horizon, the model still fails half the time, so for work you need done reliably, aim well below it and keep tests and review in the loop. The 50-minute figure is from early 2025; check the authors' later updates before quoting a current number. For agent design that raises reliability, see SWE-agent and agentic workflows.
Abstract
Despite rapid progress on AI benchmarks, the real-world meaning of benchmark performance remains unclear. To quantify the capabilities of AI systems in terms of human capabilities, we propose a new metric: 50%-task-completion time horizon. This is the time humans typically take to complete tasks that AI models can complete with 50% success rate. We first timed humans with relevant domain expertise on a combination of RE-Bench, HCAST, and 66 novel shorter tasks. On these tasks, current frontier AI models such as Claude 3.7 Sonnet have a 50% time horizon of around 50 minutes. Furthermore, frontier AI time horizon has been doubling approximately every seven months since 2019, though the trend may have accelerated in 2024. The increase in AI models' time horizons seems to be primarily driven by greater reliability and ability to adapt to mistakes, combined with better logical reasoning and tool use capabilities. We discuss the limitations of our results -- including their degree of external validity -- and the implications of increased autonomy for dangerous capabilities. If these results generalize to real-world software tasks, extrapolation of this trend predicts that within 5 years, AI systems will be capable of automating many software tasks that currently take humans a month.

