{"version":1,"asset_type":"statistical_series","data_type":"time_series","slug":"ai-autonomous-task-horizon","url":"https://apiardata.com/statistics/ai-autonomous-task-horizon","html_url":"https://apiardata.com/statistics/ai-autonomous-task-horizon/","title":"AI Autonomous Task Time Horizon — How Long Frontier Models Work Without Human Intervention (2020–2026)","description":"Tracks the 50% task time horizon of frontier AI agents on METR's HCAST benchmark: the length of task (measured by human expert completion time) at which a model succeeds roughly half the time when working autonomously. From 9 seconds for early GPT-3 agents in 2020 to 14.5 hours for Claude Opus 4.6 in February 2026, the horizon has roughly doubled every 7 months.","domain":"papers","category":"AI & Research","keywords":["AI & Research","AI Agents","AI Autonomy","AI Benchmarks","AI Safety"],"publisher":"METR (Model Evaluation & Threat Research); arXiv:2503.14499 (Kwa et al., March 2025)","frequency":"Quarterly (approx.)","geography":"Global — frontier AI systems","temporal_coverage":"2020/..","last_updated":"2026-04-01","last_updated_text":"April 2026","data_as_of":"February 2026","variable_measured":"50% task time horizon — the human expert completion time at which a frontier AI agent achieves approximately 50% task success rate when operating autonomously","measurement_technique":"Logistic-curve fitting across hundreds of software-oriented agentic tasks pre-rated by human annotators for expert completion time; 50% success-rate threshold derived from bootstrapped model fit; evaluations conducted in sandboxed Linux environments with no human intervention during task execution (METR HCAST and Time Horizon 1.1 methodology)","license":"https://apiardata.com/data-license","is_accessible_for_free":true,"sources":[{"name":"METR — Blog post: Measuring AI ability to complete long tasks (Mar 2025)","url":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/"},{"name":"METR — Interactive time-horizons dashboard","url":"https://metr.org/time-horizons/"},{"name":"METR — Time Horizon 1.1 update (Jan 2026)","url":"https://metr.org/blog/2026-1-29-time-horizon-1-1/"},{"name":"METR — GPT-5 evaluation report","url":"https://evaluations.metr.org/gpt-5-report/"},{"name":"Epoch AI — METR Time Horizons benchmark tracking","url":"https://epoch.ai/benchmarks/metr-time-horizons"}],"meta":[{"label":"Frequency","value":"Quarterly (approx.)"},{"label":"Coverage","value":"2020–2026 (8 frontier model evaluations)"},{"label":"Observations","value":"8"},{"label":"Geography","value":"Global — frontier AI systems"},{"label":"Last updated","value":"April 2026"}],"kpis":[{"label":"50% Task Horizon — Claude Opus 4.6 (Feb 2026)","value":"14.5 hours","unit":"METR Time Horizon 1.1 evaluation","trend":{"direction":"up","value":"vs. 9 seconds in 2020"}},{"label":"Historical Doubling Rate (2020–2025)","value":"~7 months","unit":"Average time to double the 50% task horizon","trend":{"direction":"up","value":"Recently accelerating to ~4 months in 2024–2025"}},{"label":"50% Task Horizon — GPT-5 (Jun 2025)","value":"2h 17m","unit":"METR direct evaluation","trend":{"direction":"up","value":"~34× increase since GPT-4 in 2023"}},{"label":"Total Increase (2020 → 2026)","value":"~4 orders of magnitude","unit":"9 seconds → 14.5 hours","trend":{"direction":"up","value":"~5,800× in absolute terms"}}],"series":[{"key":"horizon_seconds","label":"50% task time horizon (METR HCAST)"}],"observation_count":8,"observations":[{"period":"Mid-2020","horizon_seconds":9},{"period":"Mar 2023","horizon_seconds":240},{"period":"May 2024","horizon_seconds":480},{"period":"Oct 2024","horizon_seconds":2400},{"period":"Mar 2025","horizon_seconds":4500},{"period":"Jun 2025","horizon_seconds":8220},{"period":"Jul 2025","horizon_seconds":17340},{"period":"Feb 2026","horizon_seconds":52200}],"table":{"columns":[{"key":"model","label":"Model"},{"key":"org","label":"Organisation"},{"key":"releaseDate","label":"Date"},{"key":"horizon_50pct","label":"50% Task Horizon"},{"key":"horizon_notes","label":"Notes"},{"key":"source","label":"Source"}],"rows":[{"model":"Claude Opus 4.6","org":"Anthropic","releaseDate":"Feb 2026","horizon_50pct":"14.5h","horizon_notes":"METR TH1.1 task suite; point estimate with ±30–50% uncertainty","source":"METR Time Horizon 1.1"},{"model":"Claude Opus 4.5","org":"Anthropic","releaseDate":"Jul 2025","horizon_50pct":"4h 49m","horizon_notes":"Direct METR evaluation; TH1.1 suite includes more 8h+ tasks","source":"METR — LessWrong post"},{"model":"GPT-5","org":"OpenAI","releaseDate":"Jun 2025","horizon_50pct":"2h 17m","horizon_notes":"Direct METR evaluation","source":"METR GPT-5 evaluation report"},{"model":"o3 / Claude 3.7 Sonnet","org":"OpenAI / Anthropic","releaseDate":"Q1 2025","horizon_50pct":"~75 min","horizon_notes":"METR Time Horizon 1.0 estimate; extended trend","source":"METR time-horizons page"},{"model":"Claude 3.5 Sonnet","org":"Anthropic","releaseDate":"Oct 2024","horizon_50pct":"~40 min","horizon_notes":"Late-2024 frontier cohort; HCAST+SWAA benchmark","source":"METR arXiv:2503.14499"},{"model":"GPT-4o","org":"OpenAI","releaseDate":"May 2024","horizon_50pct":"~8 min","horizon_notes":"METR HCAST+SWAA evaluation","source":"METR arXiv:2503.14499"},{"model":"GPT-4","org":"OpenAI","releaseDate":"Mar 2023","horizon_50pct":"~4 min","horizon_notes":"HCAST evaluation","source":"METR arXiv:2503.14499"},{"model":"GPT-3 agents","org":"OpenAI","releaseDate":"Mid-2020","horizon_50pct":"~9s","horizon_notes":"Retrospective estimate; early agentic scaffolding","source":"METR arXiv:2503.14499"}]},"qa":[{"question":"What does the 50% task time horizon mean and how is it measured?","answer":"The 50% task time horizon is the length of task at which a frontier AI agent succeeds approximately half the time when operating autonomously without human guidance. Task length is measured in human expert completion time — the estimated time a skilled human professional would require to perform the same task from scratch — not in wall-clock AI execution time. That distinction matters: a task rated at 40 minutes of human expert time might take an AI agent hours, or might be completed much faster through parallelism; the metric captures task complexity rather than AI speed. METR derives this estimate by fitting a logistic curve to a large set of tasks, each pre-rated by human annotators for complexity, then running AI agents on those tasks in a sandboxed environment. The 50% success rate threshold is chosen because it sits in the most informative region of the logistic curve — where task difficulty and agent capability interact most visibly. The resulting metric compresses the agent's capability distribution to a single interpretable number: the task difficulty, in human-time units, where the agent is about as likely to succeed as to fail."},{"question":"Why is the y-axis displayed on a logarithmic scale?","answer":"The 50% task time horizon has grown from approximately 9 seconds in mid-2020 to 52,200 seconds (14.5 hours) by February 2026 — a factor of roughly 5,800 across six years. Plotting this trajectory on a linear axis would make the early data points (GPT-3 agents at 9 seconds, GPT-4 at 240 seconds) visually indistinguishable from zero, while compressing all the interesting variation into a near-vertical spike at the right edge. A logarithmic scale allows every doubling of the horizon to appear as an equal vertical step, which matches the underlying pattern in the data: the horizon has roughly doubled every 7 months over most of the observed period. On a log scale, a straight trend line indicates consistent exponential growth; deviations from linearity reveal acceleration or deceleration. The chart therefore does not exaggerate the trend — it is the scale that makes the empirical regularity visible without distortion."},{"question":"What is HCAST and how broad is its coverage?","answer":"HCAST (Human-Comparable Agentic Software Tasks) is a benchmark suite developed by METR that consists of hundreds of real-world software engineering and technical tasks sourced from online task platforms, professional workflows, and internal test development. Tasks span Python scripting, shell automation, repository navigation, data analysis pipelines, web scraping, bug reproduction, and similar programming-adjacent work. Agents are evaluated in a sandboxed Linux environment with access to standard tools, internet-connected resources where tasks require them, and no human interaction during the task. The benchmark is specifically designed so that each task has a verifiable, objective completion criterion — allowing binary scoring without subjective judgment. The critical scope limitation is that HCAST is predominantly a software engineering benchmark. The majority of tasks are solvable by someone with programming competence; general professional tasks requiring sustained natural-language reasoning, interpersonal coordination, visual judgment, or domain expertise outside software are largely absent from the current suite. The Time Horizon 1.1 update published in January 2026 expanded the task set with more tasks in the 4–12 hour range, which widened the measurable range but did not substantially broaden the task type distribution. Inferences about agent capability in non-technical workplace contexts should be made cautiously from this data."},{"question":"Is the doubling trend likely to continue?","answer":"The observed doubling time of approximately 7 months (compressing to roughly 4 months in the 2024–2025 period) is an empirical pattern in a small historical dataset of eight data points, not a physical law or a guaranteed trajectory. Several factors could cause the trend to slow: benchmark saturation is possible if current task sets become fully solved before longer tasks are added; the tasks that remain hardest may require qualitatively different capabilities not currently improving at the same rate; and real-world deployment constraints — authentication, approval gates, organisational context — not present in HCAST may limit practical capability gains even as benchmark scores continue to rise. The trend could also accelerate if architectural changes, improved scaffolding, or expanded tool access produce larger-than-typical capability jumps in individual model releases. METR's own published commentary frames the trend descriptively rather than predictively. No published mechanistic model explains why the doubling rate should be stable or should change at any particular threshold. Any forward projection from this chart carries substantial uncertainty and should be treated as illustrative rather than predictive."},{"question":"What does this benchmark imply for autonomous AI deployment in practice?","answer":"The benchmark measures performance in a controlled, sandboxed environment on software-oriented tasks with objective completion criteria. Autonomous task completion in real enterprise workflows typically introduces several layers of complexity absent from HCAST: authentication and access management across multiple systems; human approval gates embedded in procurement, legal, and compliance workflows; integration with legacy software not designed for API or agent access; ambiguous or evolving task specifications that require iterative human clarification; and accountability structures where errors have real organisational consequences. A model capable of completing a 14.5-hour benchmark task autonomously may still require substantial human oversight infrastructure when deployed on tasks of equivalent nominal duration in a live enterprise setting. The benchmark is most directly informative for organisations building agentic pipelines in software development, data engineering, or technical operations — where HCAST's task composition more closely mirrors actual work. For policy researchers and investors, the metric provides a quantitative proxy for the pace of capability development; it does not directly translate to workforce substitution estimates or deployment-readiness conclusions without additional empirical evidence from real deployment contexts."},{"question":"How reliable are the individual data point estimates, particularly for more recent evaluations?","answer":"Reliability varies substantially across the dataset. The mid-2020 data point is a retrospective estimate by METR researchers applied to early GPT-3-era agentic systems; it was not a contemporaneous evaluation and carries higher uncertainty than the later points. Data points from METR's March 2025 paper (arXiv:2503.14499) — covering GPT-4 through the late-2024 frontier cohort — are derived from the most systematic evaluation methodology in the dataset: logistic-curve fitting across hundreds of tasks with bootstrapped confidence intervals. The paper notes that point estimates carry roughly ±30–50% uncertainty expressed as a ratio; a stated horizon of 40 minutes plausibly covers a range from approximately 20 minutes to 60 minutes. The post-March-2025 data points (o3/Claude 3.7 Sonnet, GPT-5, Claude Opus 4.5, Claude Opus 4.6) come from METR's ongoing evaluation programme using the TH1.1 task suite, which added more tasks in the 4–12 hour range specifically to ensure coverage at higher capability levels. The expansion of the task suite means the TH1.1 estimates are not strictly methodologically identical to the earlier HCAST estimates in the 2025 paper — longer-task performance may be measured more precisely in TH1.1, but the calibration anchor differs. Epoch AI's benchmark tracking page documents available confidence intervals alongside point estimates; for quantitative modelling applications, the confidence interval data should be incorporated rather than relying solely on point estimates."}]}