AI Completes Half-Day Tasks in Minutes. Important Takeaways from METR's Evaluations of Claude's Time Horizons

AI is no longer just assisting work—it’s accelerating hours of human effort into minutes, reshaping how autonomy must be governed

Recent evaluations by of Anthropic’s newest Claude model mark an important shift in how advanced AI systems are assessed. Rather than focusing on accuracy scores or benchmark rankings, METR evaluates how much real-world work an AI system can reliably perform over time. This approach is especially relevant for organizations moving from pilots to real operational use of AI.

One of METR’s most cited metrics is the model’s “time horizon.” This measures the length of a task—defined by how long it typically takes a skilled human to complete—that an AI system can successfully finish about 50% of the time without major intervention. For the latest Claude Opus model, METR estimates a 50% time horizon of approximately 4 hours and 49 minutes. Crucially, this does not mean the AI runs for nearly five hours. In practice, tasks in this category typically take Claude minutes to tens of minutes of wall-clock time, depending on task complexity and tool use. The metric reflects human labor equivalence, not execution duration.

Other METR results reinforce this pattern of capability compression. In METR’s General Autonomous Capabilities (GAC) evaluations, earlier Claude Sonnet models successfully completed around 50% of tasks that required human experts roughly 55 minutes to finish. In separate AI research and development–style tasks (RE-Bench), Claude approached median human expert performance under controlled conditions. Together, these metrics show that modern AI systems are no longer limited to short, single-step outputs—they can handle sustained, multi-step workstreams.

For business leaders, the implication is straightforward: hours of expert human effort can now be compressed into minutes of AI execution. This creates clear productivity, cost, and speed advantages—but it also changes the risk profile. An AI system that can independently complete hours’ worth of work can just as quickly propagate errors, reinforce flawed assumptions, or act outside intended boundaries if oversight and controls are insufficient.

The takeaway is not alarm, but intentional adoption. METR’s findings highlight why organizations must think carefully about where AI is granted autonomy, how its outputs are reviewed, and when human intervention is required. As AI systems like Claude cross thresholds of longer independent operation, governance must evolve accordingly. METR’s evaluation makes one thing clear: AI risk is no longer just about what a model knows—it is about how much work it can do, how fast, and with how little supervision.

FYI: While this discussion focuses on METR’s autonomy and “time horizon” evaluations, METR also conducts benchmarking across other risk-relevant areas, including cybersecurity capabilities, dual-use scientific reasoning, deception and persuasion, and AI-assisted research and development. These assessments are often more restricted or qualitative, but they serve a complementary purpose: identifying where AI systems may meaningfully lower the barrier to harm or amplify misuse if deployed without safeguards. Together, these benchmarking streams help organizations understand not just how productive an AI system can be, but where and why additional controls may be required.

Bring Venra Into Your Transformation

At Venra Labs, we help organizations introduce technology the right way — with clean data, clear processes, responsible governance, and people-centered change woven into every step. Whether you're rolling out AI, automation, or modern data workflows, we ensure your teams understand the tools, trust them, and feel empowered using them.

If your organization is preparing for a technology rollout and wants adoption from day one, let’s partner to make your transformation smooth, safe, and successful.

👉 Book a readiness call with Venra Labs