AI Autonomous Task Time Horizon — How Long Frontier Models Work Without Human Intervention (2020–2026)

In mid-2020, an AI agent built on an early GPT-3-class model could autonomously complete tasks requiring roughly 9 seconds of equivalent human expert effort at the 50% success rate threshold. By April 2026, an early preview of Claude Mythos reached a 50% task horizon of approximately 17.4 hours on METR's Time Horizon 1.1 (TH1.1) benchmark — provisionally, since that figure sits above the roughly 16-hour range where METR's current task suite can produce a reliable estimate. That progression — from 9 seconds to 62,687 seconds — spans nearly four orders of magnitude (roughly 7,000×) in under six years and represents the most systematically measured trajectory of AI agentic capability currently available in published research.

The metric behind this chart is METR's task time horizon, introduced in the March 2025 paper "Measuring AI Ability to Complete Long Tasks" (Kwa et al., arXiv:2503.14499) and since re-baselined under the Time Horizon 1.1 methodology. The 50% task time horizon is defined as the human expert completion time of the task at which a frontier agent succeeds approximately half the time when working entirely autonomously. "Human expert completion time" is the critical unit: it measures the difficulty of the task as a proxy, not how long the AI takes to execute. Every point on this page, from GPT-4 in March 2023 through Claude Mythos Preview in April 2026, is now drawn from the same TH1.1 dataset, resolving a partial methodological mismatch that existed in earlier versions of this chart, when historical points came from the original HCAST-only evaluation and newer points came from TH1.1.

METR's own curve fit puts the doubling time for the 50% task time horizon at roughly 6.2 months across the full 2020–2026 history, accelerating to roughly 4.2 months when the fit is restricted to 2023 onward — a rate that, if sustained, would imply day-length autonomous task capability within the next one to two years. Whether that rate is sustained remains empirically open, and METR's own task suite already struggles to measure reliably above roughly 16 hours.

The growth pattern is most clearly visible on the logarithmic scale used in the chart above. On a log axis, each doubling of the task horizon appears as an equal vertical increment. The trajectory from GPT-4 (approximately 4 minutes, March 2023) to GPT-4o (approximately 7 minutes, May 2024) represents a modest single-year gain; from GPT-4o to Claude 3.5 Sonnet (approximately 21 minutes, October 2024) is roughly one and a half additional doublings across five months. The frontier has kept moving since: Claude Opus 4.6 reached roughly 12 hours (11h 59m) by February 2026 — the same week OpenAI's GPT-5.3-Codex reached roughly 5h 50m, illustrating that different labs' frontier releases in the same month can sit far apart on this metric. An early preview of Claude Mythos then reached a provisional ~17.4 hours by April 2026, the highest figure recorded and the first point to exceed METR's stated reliable-measurement ceiling. As of September 2026, METR had not yet published time horizons for several newer releases, including Claude Opus 4.7, GPT-5.5 and Grok 4.3, so Claude Mythos Preview remains the latest point in the dataset.

Two scope limitations are essential context for professional use of this data. First, the underlying task suite (HCAST, extended under Time Horizon 1.1) is predominantly a software engineering benchmark. Most tasks involve Python scripting, shell automation, data pipelines, repository navigation, and related programming-adjacent work in a sandboxed Linux environment. The benchmark does not systematically cover the sustained natural-language reasoning, interpersonal coordination, and domain-specific judgment that characterise many professional workflows. A multi-hour task horizon on this benchmark does not imply equivalent autonomous capability across general office work. Second, point estimates carry approximately ±30–50% uncertainty as ratios, and the most recent point is explicitly flagged by METR as provisional: its own published estimate for Claude Mythos Preview is "at least 16 hours," with a 95% confidence interval spanning 8.5 to 55 hours, considerably wider than the single point estimate plotted on this chart. In March 2026, METR also corrected a regularisation error in its estimation pipeline, which revised several point estimates — including Claude Opus 4.6's — materially downward from what had previously been reported; the figures on this page reflect that correction.

The mid-2020 data point deserves a separate note: it is METR's earliest Time Horizon 1.1 baseline, applied retrospectively to an early GPT-3-class model, not a contemporaneous evaluation. Systematic task-horizon methodology did not exist in 2020; the roughly 9-second estimate is derived by back-fitting the observed trend and applying it to documented early agent performance. It should be treated as an indicative anchor rather than a precision measurement.

For policy researchers and investors, the practical signal in this dataset is primarily about pace of capability development rather than deployment readiness. Benchmark performance in a controlled sandbox environment and autonomous operation in real enterprise workflows involve meaningfully different conditions: live enterprise settings typically include authentication and access management, embedded human approval gates, integration with legacy systems not designed for agent access, and accountability structures where errors carry organisational consequences. The gap between benchmark capability and deployable autonomy is not constant, and measuring it requires empirical evidence from real deployment contexts — which the HCAST/TH1.1 benchmark alone does not provide.

Key methodological notes for quantitative use of this dataset:

• Task time unit: "Human expert completion time" is a difficulty proxy rated by annotators, not AI execution time. The two can differ substantially.
• Logistic curve fitting: The 50% horizon is a model-derived estimate, not a directly observed value. Point estimates carry ±30–50% uncertainty as ratios.
• Benchmark scope: HCAST and TH1.1 are software-engineering-heavy. Performance on general professional tasks may differ substantially.
• Single consistent methodology: Every point on this page is drawn from METR's Time Horizon 1.1 dataset, applied consistently from GPT-4 (2023) through Claude Mythos Preview (2026), replacing the mixed HCAST/TH1.1 sourcing used in earlier versions of this chart.
• Measurement ceiling: METR's current task suite reliably measures horizons up to roughly 16 hours. The Claude Mythos Preview point (~17.4 hours) exceeds that ceiling and is provisional; METR's own published range for it is 8.5 hours to 55 hours (95% confidence interval).
• Regularisation correction: METR corrected an error in its estimation pipeline in March 2026, revising some point estimates — including Claude Opus 4.6's — downward from previously reported levels.
• Retrospective baseline: The mid-2020 estimate is a back-fitted retrospective, not a contemporaneous evaluation.
• Trend extrapolation: The observed doubling rate is a historical pattern in seventeen data points. It carries no mechanistic guarantee of continuation, acceleration, or saturation.