METR's groundbreaking research, published in March 2025, introduces a novel metric for measuring AI progress: the 50% task completion time horizon. This measures the duration of tasks (calibrated to human completion times) that AI can solve with 50% reliability.

The historical progression is striking:

This progression reveals a clear exponential trend, with capabilities doubling approximately every 7 months—significantly faster than traditional computing advances like Moore's Law (which doubles every 2 years).

METR's findings are based on the Human-Calibrated Autonomy for Software Tasks (HCAST) benchmark, which evaluates AI performance on diverse software engineering tasks ranging from 1 second to 16 hours of human completion time.

The methodology involved:

1. Human calibration: Timing skilled humans on a combination of benchmarks (HCAST, RE-Bench, and 66 novel shorter tasks)
2. Diverse task domains: Software engineering, machine learning, cybersecurity, and general reasoning
3. Controlled evaluation: Testing AI models under similar conditions to humans
4. Success rate analysis: Plotting success rates against human completion times to derive the 50% threshold

This approach provides a more realistic assessment of AI capabilities than traditional benchmarks, which often focus on narrow skills without considering task complexity or time requirements.

Current frontier models like Claude 3.7 Sonnet demonstrate a 50% time horizon of approximately 50 minutes. This means they can complete tasks that would take skilled humans about 50 minutes with 50% reliability.

The success rate varies significantly by task duration:

• ~100% success on tasks taking humans less than 4 minutes
• ~50% success on tasks taking humans around 50 minutes
• <10% success on tasks taking humans more than 4 hours

This pattern reveals both the impressive progress of AI and its current limitations. While AI excels at shorter, knowledge-intensive tasks, its reliability drops significantly for longer, more complex problems that require sustained reasoning, planning, and error correction.

# Example of a task at the current frontier (50-minute human time)
def optimize_database_query(query_string, schema, sample_data):
    """
    Analyze and optimize a complex SQL query for performance
    while maintaining identical results.
    
    1. Parse and understand the original query structure
    2. Identify performance bottlenecks (missing indexes, inefficient joins)
    3. Rewrite the query with optimizations
    4. Verify identical results using sample data
    5. Explain optimization rationale
    """
    # AI can complete this type of task with ~50% reliability

Extrapolating the 7-month doubling trend yields remarkable projections for AI capabilities:

These projections suggest that within 3-4 years, AI could autonomously handle tasks that currently require teams of human developers working for days or weeks.

Transformative Benefits

• Productivity revolution: Automation of increasingly complex knowledge work
• Accelerated innovation: Faster software development and research cycles
• Democratized expertise: Access to AI capabilities that match human experts
• New business models: Services built around AI-human collaboration

Significant Challenges

• Labor market disruption: Potential displacement of knowledge workers
• Skill obsolescence: Rapid changes in valuable human capabilities
• Safety concerns: Risks from increasingly autonomous systems
• Regulatory gaps: Need for frameworks to manage powerful AI systems

METR's findings align with other frameworks for measuring AI progress:

• Richard Ngo's t-AGI framework: Compares AI to time-limited human experts
• Bio Anchors approach: Projects AI development timelines based on computational requirements
• Scaling laws: Predicts capability improvements based on compute, data, and parameter scaling

The 7-month doubling time provides a concrete, empirically-grounded metric that complements these theoretical approaches, offering a practical way to forecast AI capabilities across diverse domains.

METR's research provides compelling evidence that AI capabilities are advancing at an exponential rate, with task completion length doubling every 7 months. This pattern suggests we are on the cusp of a profound transformation in knowledge work, with AI potentially handling week-long tasks by 2028 and month-long projects by 2029.

This rapid progression demands strategic foresight from business leaders, policymakers, and technologists. Rather than viewing AI as a static technology, organizations must prepare for a dynamic landscape where capabilities expand predictably but dramatically over time.

The 7-month doubling metric offers a valuable tool for navigating this future—providing a concrete timeline for capability development that can inform investment decisions, workforce planning, and regulatory approaches. By understanding this exponential trend, stakeholders can better prepare for both the transformative benefits and significant challenges of increasingly capable AI systems.