METR finds frontier AI software-task horizons doubling about every seven months

On March 19, 2025, METR published a method for measuring model-agent capability using the estimated time a relevant human expert would need to complete a task. The original suite combined 97 HCAST software tasks, seven RE-Bench machine-learning research tasks, and 66 short Software Atomic Actions, covering 170 tasks in total.

For each agent, METR modelled success probability as a function of the logarithm of human completion time. Its 50% task-completion time horizon is the human task duration at which the fitted model predicts that the agent will succeed half the time. In METR’s original March 2025 disclosure and version, Claude 3.7 Sonnet’s estimate was approximately 50 minutes, while the fitted frontier trend implied a doubling roughly every seven months from 2019 to early 2025.

The peer-reviewed NeurIPS 2025 version later evaluated agents based on 12 frontier models and reported o3 at roughly a 110-minute 50% horizon, while retaining the approximately seven-month long-run doubling trend. METR later expanded the suite from 170 to 228 tasks in Time Horizon 1.1, increasing the number of tasks estimated to take humans at least eight hours from 14 to 31.

METR’s original March 2025 disclosure found a strong relationship between human task duration and model-agent success across the evaluated software and research tasks. It reported an approximately exponential increase in the frontier 50% horizon between 2019 and early 2025 and found a similar growth trend at the stricter 80% success threshold, although the 80% horizons were approximately five times shorter. The later peer-reviewed NeurIPS 2025 evaluation retained the approximately seven-month long-run trend while reporting its results for 12 frontier models, including o3 at roughly a 110-minute 50% horizon.

The later TH1.1 evaluation preserved the stitched long-run result: approximately 196.5 days per doubling, effectively the same as the original TH1 estimate of 195.8 days. Shorter-window estimates were faster—approximately 130.8 days from 2023 onward and 88.6 days from 2024 onward under TH1.1—but METR found that these recent slopes were more sensitive to task composition and the set of models included.

A separate ICML 2026 paper, BRIDGE, used a two-parameter item-response model and performance data from METR, SWE-bench Verified, MLE-bench, GDPval, and Cybench. Its authors independently recovered an exponential frontier trend with a roughly six-month doubling time. This supports the broad scaling result, although BRIDGE still used METR’s human-time annotations to calibrate its latent difficulty scale and therefore is not fully independent of the original dataset.

METR has published the analysis pipeline and associated data used to calculate the time horizons, allowing the statistical analysis and generated figures to be inspected and rerun.

A time horizon is not the amount of wall-clock time an AI can continuously operate without intervention. It is the amount of human expert task time associated with tasks the evaluated agent is predicted to complete at a specified success rate. Agents often finish successful tasks substantially faster than humans.

The result measures complete agent configurations rather than bare language models. Performance depends on the model, scaffold, prompts, tools, context management, resource limits, and elicitation effort. Most models in the original study used METR’s modular-public scaffold, while some models used modified setups because of differences in tool-use behaviour.

The evaluated distribution was concentrated on automatically gradable, self-contained software and research tasks. These tasks generally require less organizational context, collaboration, ambiguous goal interpretation, and accumulated codebase knowledge than normal long-running workplace projects. The original paper found that agents performed worse on “messier” tasks even after controlling for human task duration.

Human completion times are also an imperfect difficulty measure. Baselines were collected from people with different skills, some longer tasks used researcher estimates rather than measured completions, and successful runs were used to calculate task length. METR’s expanded TH1.1 suite had measured human baselines for only five of its 31 tasks estimated at eight hours or longer.

Later analysis found that recent frontier estimates are increasingly sensitive to task selection, benchmark saturation, regularization, curve-fitting assumptions, and uncertainty in human task lengths. A corrected regularization mistake reduced some recent 50% horizon estimates by as much as 20%, while reasonable alternative modelling choices could shift recent estimates further. These issues affect precise model values more than the broad multi-year trend.

The durable result is that model agents improved extremely rapidly on METR’s human-calibrated software and research tasks across multiple model generations, and that human task duration provided a useful common scale for comparing systems whose raw benchmark scores were otherwise difficult to relate.

The entry should not be read as evidence that a model with a one-hour horizon can perform every one-hour professional task, work independently for an hour, replace an hour of labour without supervision, or reliably complete tasks below that horizon. It also does not establish that the faster post-2024 slope will continue or that month-long general workplace autonomy will arrive on the paper’s extrapolated schedule.