Kwa, West, Becker, Deng, Garcia et al. (METR) — 2026

Measuring AI in Human Time

Forget benchmark percentages. The real question: how long would a task take a human expert? RE-Bench measures AI capability as a time horizon — and it doubles every 7 months.

Prerequisites: AI benchmarking basics + logistic regression + exponential growth
10
Chapters
4+
Simulations

Chapter 0: The Problem

Imagine someone tells you: "GPT-4 scores 87% on SWE-bench." What does that mean? Can it write a web app? Fix a production bug? Refactor a codebase? You have no idea. The number is meaningless without a human reference frame.

This is the fundamental problem with AI benchmarks. They tell you about relative ordering between models, but nothing about absolute capability in human terms. Benchmarks saturate within months. GLUE was solved in two years. SuperGLUE lasted barely longer. MMLU went from 50% to 90% in under three years. Each time, a new, harder benchmark appears — and the cycle repeats.

The deeper issue: We cannot compare models across different benchmarks. What does it mean that GPT-2 scored 30% on LAMBADA while o3 scores 87% on SWE-bench Verified? These are different yardsticks measured at different times. There is no common unit of capability.

METR (Model Evaluation & Threat Research) noticed this gap and asked a simple question: what if we measured AI capability not in percentage points, but in human time?

The Benchmark Saturation Problem

Each benchmark saturates, then gets replaced. Click through the timeline to see the pattern.

GLUE (2018)

The pattern is always the same: a new benchmark appears, models struggle, rapid improvement follows, the benchmark saturates, everyone moves on. But each benchmark measures something different, so you cannot draw a single trend line through them. You cannot answer the question that actually matters: how capable are AI systems right now, in terms a human can understand?

Why are individual benchmark scores insufficient for tracking AI progress over time?

Chapter 1: The Key Insight

Here is the core idea, in one sentence:

Measure AI capability in human time. Instead of asking "what percentage of tasks can the model solve?", ask "how long would the tasks it can solve take a human expert?" An AI with a 2-hour time horizon can reliably do what a skilled human does in 2 hours.

This is inspired by human psychometrics. When you take an IQ test, the score is calibrated against human performance — it tells you where you stand relative to people, not just how many questions you got right. METR does the same thing for AI: calibrate against timed human experts.

The key metric is the 50%-task-completion time horizon: the duration of tasks that the AI can complete with 50% success rate. If an AI's time horizon is 110 minutes, it means:

Why 50%? It is a natural midpoint. Below 50%, the model is unreliable. Above 50%, it is fairly dependable. The 50% threshold gives you the frontier of what the model can barely do.

Why this works across time: Human time is a universal currency. A task that takes a human 30 minutes in 2020 still takes a human 30 minutes in 2025. So when we say GPT-2 (2019) had a 2-second time horizon and o3 (2025) has a 110-minute time horizon, we are comparing them on the same scale, even though they never took the same benchmark.

The beauty of this framing: it turns AI progress into a single, interpretable number that grows over time. And that number, it turns out, follows a strikingly clean exponential trend.

What does "50%-task-completion time horizon of 110 minutes" mean concretely?

Chapter 2: The Time Horizon Metric

Let us formalize how the time horizon is actually computed. You need three ingredients: a set of tasks, human completion times for those tasks, and AI success rates on those tasks.

Step 1: Binary success

Each AI attempt on a task is converted to a binary outcome: success or failure. Some tasks are naturally binary ("did the shell script run correctly?"). Continuously scored tasks use a threshold based on human performance — if a human working for 7-9 hours achieves a certain score, the AI must match it to count as "success."

Step 2: Task timing

Each task has a human time rating ttask, defined as the geometric mean time of successful human baselines. This is how long a skilled domain expert takes to complete the task. Tasks range from 1 second (trivial shell commands) to 8+ hours (complex ML research engineering).

Step 3: Logistic regression

Now we fit a model that predicts success probability as a function of task length:

psuccess(agent, task) = σ((log hagent − log ttask) · βagent)

where σ is the sigmoid function, hagent is the agent's time horizon (what we are solving for), and βagent controls how steeply success drops off with task length. When ttask = hagent, the argument to σ is zero, so psuccess = 0.5. That is exactly the definition: hagent is the task duration at which success probability is 50%.

The Logistic Fit

Drag the time horizon slider to see how the success curve shifts. At hagent, success is exactly 50%.

hagent110 min
Why log-space? The regression is in log-time, not linear time. This is because task difficulty scales roughly logarithmically with duration — a 10-minute task is not 5x harder than a 2-minute task. The exponential spread of durations (1 second to 8 hours) means log-space produces a much better fit.

This approach is inspired by Item Response Theory (IRT) from psychometrics — the same framework used to calibrate standardized tests like the GRE. In IRT, each test item has a difficulty parameter and each test-taker has an ability parameter. Here, task duration replaces item difficulty, and the time horizon replaces test-taker ability.

In the logistic regression p = σ((log h − log t) · β), what happens when the task duration t equals the time horizon h?

Chapter 3: Task Design

The task suite must cover a wide range of difficulty. Too easy and you cannot distinguish frontier models. Too hard and older models score zero across the board. METR combines three datasets to span nearly the entire difficulty spectrum:

1. SWAA: Software Atomic Actions (66 tasks)

These are the shortest tasks — single-step operations that a software developer does hundreds of times a day. They take 1 second to 30 seconds for a human:

These tasks are crucial because they let the researchers measure pre-2023 models like GPT-2 and GPT-3, which cannot do anything more complex.

2. HCAST: Human-Calibrated Autonomy Software Tasks (97 tasks)

The middle range — real software tasks from 1 minute to 30 hours:

3. RE-Bench: Research Engineering Benchmark (7 tasks)

The hardest tasks — all rated at 8 hours. These are genuine ML research engineering problems: optimize a neural network for efficiency, implement a complex algorithm from a paper description, debug a distributed training pipeline.

Task Difficulty Spectrum

170 tasks spanning 5 orders of magnitude in human time. Each dot is a task. Hover to see examples.

Why three datasets? No single benchmark covers the full range. SWAA handles the trivial end (seconds), HCAST covers the middle (minutes to hours), and RE-Bench covers the hard end (full workday). Together, they span from 1-second tasks to 8-hour tasks — roughly 5 orders of magnitude on a log scale. This is essential for fitting the logistic model across many different model generations.

In total: 170 tasks, all automatically scored (binary or threshold-based), covering software engineering, ML research, and cybersecurity domains.

Why does the task suite need to span from 1-second to 8-hour tasks?

Chapter 4: The Doubling Trend

This is the headline result. When you plot each model's 50% time horizon against its release date, the points fall on a clean exponential line.

AI time horizon doubles every ~7 months. GPT-2 (June 2019): ~2 seconds. GPT-3 (November 2020): ~3 minutes. GPT-4 (March 2023): ~30 minutes. o3 (January 2025): ~110 minutes. Each generation roughly doubles the duration of tasks AI can handle.
The Exponential Growth of AI Time Horizon

Each point is a frontier model. The line is an exponential fit. Drag the extrapolation slider to see future predictions.

Extrapolate to2025

The actual data points from the paper, in chronological order:

GPT-2 (Jun 2019)
~2 seconds — can identify which file is a shell script
GPT-3 (Nov 2020)
~3 minutes — can do simple Wikipedia research tasks
GPT-3.5 (Mar 2022)
~13 minutes — can fix simple bugs in config files
GPT-4 (Mar 2023)
~33 minutes — can write data transformation scripts
Claude 3.5 Sonnet (Jun 2024)
~60 minutes — can tackle real standalone software tasks
o3 (Jan 2025)
~110 minutes — succeeds at some 4+ hour tasks

The regression is done in log-space: log(time horizon) = α + β · release_date. The fit gives a doubling time of 207 days (95% CI: 166-240 days) — roughly 7 months. The confidence interval comes from a three-level hierarchical bootstrap over task families, tasks within families, and individual runs.

One notable outlier: o3 sits above the trend line (p = 0.006), suggesting the trend may have accelerated in late 2024 and early 2025. The 2024-2025 only trend shows a doubling time closer to 5 months.

What is the approximate doubling time for AI task-completion time horizon, and what does this mean concretely?

Chapter 5: What Drives Improvement

A number going up is interesting. Understanding why it goes up is useful. METR manually inspected 31 failed runs from GPT-4 (November 2023) and 32 failed runs from o1 (September 2024) to categorize what went wrong.

The failure categories

Repeating failed actions
GPT-4: 12/31 failures (39%) → o1: 2/32 (6%). Dramatic improvement. Older models would bash their head against the same wall repeatedly. Newer models recognize and adapt.
Premature task abandonment
GPT-4: 8/31 (26%) → o1: 16/32 (50%). Interestingly, this got worse — but for a good reason. o1 tackles harder tasks where giving up is sometimes the right call. Its failures come from harder territory.
Incorrect reasoning / mental math
GPT-4: 6/31 (19%) → o1: 7/32 (22%). Still a persistent weakness, though o1's errors occur on significantly harder problems.
Poor planning / tool choice
GPT-4: 4/31 (13%) → o1: 6/32 (19%). Both models sometimes pick the wrong approach entirely.
The big pattern: Improvement is driven more by reliability than by raw capability. The single largest improvement from GPT-4 to o1 is the drop in "repeating failed actions" — from 39% to 6%. This is not about knowing more; it is about recovering from errors. Tool use, self-awareness, and adaptability matter more than pure knowledge.

The "messiness" factor

METR also scored tasks on 16 "messiness" factors — things like resource constraints, dynamic environments, lack of clear feedback, and requirement for high reliability. These factors represent the gap between benchmarks and real-world work.

Models perform worse on messier tasks, as expected. But the key finding: the growth trend is similar for both clean and messy tasks. The time horizon doubles at roughly the same rate regardless of messiness. This is cautiously encouraging for external validity — the trend does not appear to be driven solely by improvements on artificially clean benchmark tasks.

What was the single largest improvement in failure modes from GPT-4 to o1?

Chapter 6: Human Baselines

The whole methodology rests on one thing: accurate human timing. If your human baselines are wrong, your time horizons are wrong. METR invested heavily here.

Who are the baseliners?

Not random Mechanical Turk workers. These are skilled professionals in software engineering, ML, and cybersecurity. Most attended world top-100 universities. Average of ~5 years of relevant experience. They are domain experts but do not have task-specific context — they know the field but have not seen the specific codebase or problem before. Think: a strong senior engineer dropped into an unfamiliar project.

Scale of baselining

Over 800 human baselines totaling 2,529 hours of work. That is over 100 person-days of skilled engineers solving problems while being timed. Broken down:

The variance problem

Humans vary enormously. On the same task, one expert might take 30 minutes and another 3 hours. METR uses the geometric mean of successful completion times as the task's duration rating. The geometric mean is more robust to outliers than the arithmetic mean — it downweights the occasional expert who takes 10x longer because they went down a rabbit hole.

An important subtlety: The choice of human population matters enormously. On METR's internal pull request tasks, contractors took 5-18x longer to fix issues than repository maintainers. If you baseline against maintainers, AI looks worse. Against contractors, AI looks better. The time horizon is always relative to a reference human population. METR uses contractors/domain experts without task-specific context — the AI's natural comparison group, since it too has no prior context on the repo.
Human Time Variance

Simulated distribution of human completion times for a 1-hour task. The geometric mean (orange) is the task's official duration.

Why does METR use the geometric mean rather than the arithmetic mean for human completion times?

Chapter 7: Results

Let us look at the full model lineup. METR evaluated 12 frontier models (and 4 near-frontier) released between 2019 and 2025, running 8 attempts per model-task pair.

Overall success rates

When you simply average success rate across all 170 tasks (weighted by inverse square root of task family size to reduce the influence of large families):

Model Success vs. Task Duration

Success rate drops exponentially with task duration (R² = 0.83). Toggle between models.

o3 (2025)

The success-vs-length relationship

There is a strong negative correlation between task length and model success rate. The regression y = -0.07x + 0.66 (where x is log human time) gives R² = 0.83. In plain terms: for every 10x increase in task duration, success rate drops about 7 percentage points. This is remarkably consistent across models — the curve just shifts right as models improve.

The 50% vs 80% gap

The 80% time horizon (tasks the model completes 80% of the time) is 4-6x shorter than the 50% horizon. For o3: 50% horizon is ~110 minutes, but 80% horizon is only ~20 minutes. This gap reveals a key limitation: even frontier models that sometimes succeed at difficult tasks cannot reliably perform tasks of moderate length. Reliability lags capability significantly.

The reliability gap: o3 can sometimes do 4-hour tasks. But it can only reliably (80%) do 20-minute tasks. This 5x gap between "can sometimes do" and "can reliably do" is a major factor in real-world deployment. A coding assistant that succeeds 50% of the time on hour-long tasks is useful. One that succeeds 80% of the time is transformative.
What does the 5x gap between 50% and 80% time horizons tell us about current AI capabilities?

Chapter 8: Extrapolation & Implications

If the doubling trend continues, where does it lead? The paper extrapolates toward a critical threshold: one month (167 working hours) of human time.

Why one month matters

One month is approximately the onboarding period for a new hire. An AI that can do one-month tasks can:

The extrapolation

At the 2019-2025 doubling rate (7 months), the 50% time horizon reaches 167 hours between mid-2028 and mid-2031 (80% confidence interval, central estimate mid-2029). If the accelerated 2024-2025 trend continues (doubling every ~5 months), it could arrive as early as late 2026 or 2027.

Extrapolation to One-Month AI

The horizontal line is 167 working hours (1 month). Toggle between the full trend and the accelerated 2024-2025 trend.

2019-2025 trend (7 mo doubling)

Caveats and limitations

The paper is careful to flag several reasons the extrapolation might be too aggressive:

  1. External validity: Benchmark tasks are cleaner than real-world tasks. They have automatic scoring, clear specifications, and no dynamic environments. Real-world software work is messier.
  2. Task distribution: The suite covers software engineering and ML research. Generalizing to all knowledge work is uncertain.
  3. Trend breaks: Exponential trends rarely continue indefinitely. Compute scaling may slow. Algorithmic progress may plateau.
  4. The messiness gap: Real tasks involve resource constraints, other agents (humans), ambiguous specifications, and consequences for failure. None of these are well-represented in the benchmark.
The sensitivity analysis: METR tested robustness against many perturbations — bootstrapping over tasks, runs, and models; varying regression hyperparameters; adding noise to baseline times. The 80% CI width on the "1-month AI" date is about 2 years. The uncertainty is dominated by external validity concerns and potential future trend changes, not statistical noise.

Dangerous capabilities

The paper discusses implications for AI safety directly. An AI with a month-long time horizon could potentially: develop novel malware, conduct sophisticated social engineering campaigns, or autonomously discover and exploit zero-day vulnerabilities. METR argues this makes capability measurement critical for safety — you cannot set appropriate guardrails if you do not know where the frontier is.

According to METR's extrapolation, when might AI reach a one-month (167 working hours) time horizon?

Chapter 9: Connections

What this paper changes

RE-Bench / METR's time horizon paper does something benchmarks rarely do: it gives us a common currency for measuring AI progress over time. Instead of comparing apples (GLUE scores) to oranges (SWE-bench percentages), we can now say "AI capability doubles every 7 months" — a single number that spans from GPT-2 to o3 and beyond.

Related work

SWE-bench Verified
The industry standard for coding agent evaluation. METR validates their trend on SWE-bench and finds an even steeper exponential — doubling every ~70 days (but possibly inflated by annotation methodology).
Agent Evaluation Survey
The first comprehensive survey of LLM agent evaluation methods. RE-Bench's time-horizon metric fits into the broader framework of capability measurement discussed in that survey.
Item Response Theory
The psychometric framework that inspired METR's logistic regression approach. IRT has been calibrating human test-takers for decades — RE-Bench adapts it for AI agents.
Compute Scaling Laws
Kaplan et al. showed loss scales as a power law with compute. RE-Bench shows that task-completion capability — a more practically meaningful metric — scales exponentially with time.
AI Safety & Dangerous Capabilities
Phuong et al. (2024) evaluated frontier models for dangerous capabilities. RE-Bench provides a quantitative framework for predicting when those capabilities reach concerning levels.

Cheat sheet

Core metric
50%-task-completion time horizon: the human-time duration of tasks the AI solves 50% of the time
Key equation
psuccess = σ((log h − log t) · β), fit via logistic regression over tasks and runs
Headline result
o3 has a 110-minute time horizon; doubling every 207 days (95% CI: 166-240)
What drives it
Reliability > raw capability. Error recovery, tool use, self-awareness in task execution
Extrapolation
1-month AI by mid-2028 to mid-2031, possibly sooner if 2024-2025 acceleration continues
What is the key advantage of the time-horizon metric over traditional benchmark scores?