AI Doubling Rate¶
The length of tasks (measured by how long they take human professionals) that generalist frontier model agents can complete autonomously with 50% reliability has been doubling approximately every seven months for the last six years. The shaded region represents 95% Confidence Interval calculated by hierarchical bootstrap over task families, tasks, and task attempts.
Interactive Features¶
- Y-Axis Scale: Toggle between Linear and Logarithmic views to see the exponential growth pattern
- Success Rate: Switch between 50% and 80% success probability metrics
- Tooltips: Hover over data points to see model details
METR Benchmark Data¶
The following table shows AI model performance on the METR-Horizon-v1 benchmark, measuring the task horizon (in minutes) that models can complete with 50% success rate.
| Model | Release Date | Task Horizon (50%) |
|---|---|---|
| GPT-2 | 2019-02-14 | 2.4 sec |
| davinci-002 | 2020-05-28 | 8.9 sec |
| GPT-3.5 | 2022-03-15 | 36.3 sec |
| GPT-4 | 2023-03-14 | 5.4 min |
| GPT-4 Turbo | 2023-11-06 | 8.5 min |
| GPT-4 (Jan) | 2024-01-25 | 5.4 min |
| Claude 3 Opus | 2024-03-04 | 6.4 min |
| GPT-4 Turbo (Apr) | 2024-04-09 | 6.6 min |
| GPT-4o | 2024-05-13 | 9.2 min |
| Qwen 2 72B | 2024-06-07 | 2.2 min |
| Claude 3.5 Sonnet | 2024-06-20 | 18.7 min |
| o1-preview | 2024-09-12 | 22.0 min |
| Qwen 2.5 72B | 2024-09-19 | 5.2 min |
| Claude 3.5 Sonnet v2 | 2024-10-22 | 29.6 min |
| o1 | 2024-12-05 | 41.1 min |
| DeepSeek V3 | 2024-12-26 | 18.5 min |
| DeepSeek R1 | 2025-01-20 | 26.9 min |
| Claude 3.7 Sonnet | 2025-02-24 | 56.1 min |
| DeepSeek V3 (Mar) | 2025-03-24 | 23.3 min |
| o3 | 2025-04-16 | 1.6 hrs |
| o4-mini | 2025-04-16 | 1.3 hrs |
| Claude 4 Opus | 2025-05-22 | 1.4 hrs |
| Claude 4 Sonnet | 2025-05-22 | 1.2 hrs |
| DeepSeek R1 (May) | 2025-05-28 | 32.2 min |
| Gemini 2.5 Pro | 2025-06-05 | 39.5 min |
| Grok 4 | 2025-07-09 | 1.8 hrs |
| Claude 4.1 Opus | 2025-08-05 | 1.9 hrs |
| GPT-5 | 2025-08-07 | 2.3 hrs |
| Claude Sonnet 4.5 | 2025-09-29 | 2.0 hrs |
| GPT-5.1 Codex | 2025-11-19 | 2.9 hrs |
| Claude Opus 4.5 | 2025-11-24 | 4.8 hrs |
AI's Ability to Handle Long Tasks¶
Summary of the METR Research
Why This Matters¶
As artificial intelligence (AI) becomes more advanced, it's not just about answering trivia questions or writing short emails anymore. A key question now is: Can AI complete long, complex tasks the way humans can—like writing software, planning events, or conducting research?
The METR team has developed a new, easy-to-understand way to measure this:
How long a task (in human time) can today's AI complete successfully?
What Did They Measure?¶
- METR looked at 170 real-world tasks like fixing software bugs, writing reports, or planning multi-step actions.
- Each task was rated by how long it typically takes a skilled human to do it—from just a few minutes to several hours.
- Then they tested how well top AI systems performed those same tasks.
What They Found¶
- Today's best AI systems (like OpenAI's and Anthropic's) can reliably complete tasks that take up to about 5 hours of human effort.
- For very short tasks (under 5 minutes), AI is nearly perfect.
- But as tasks get longer and more complex—especially past 8 hours—AI still struggles.
- Most importantly: the ability of AI to complete longer tasks is doubling roughly every 7 months.
Why This Trend Is Big News¶
If the current pace continues:
- In 2–3 years, AI may handle tasks that take a human a full week or more.
- In 5 years, it may independently manage projects that currently take a team of people a month.
This means AI could soon:
- Write complete software products
- Research and draft business strategies
- Conduct customer support or internal reporting workflows end-to-end
Things to Keep in Mind¶
- A 50% success rate isn't perfect. AI may still make mistakes or need supervision.
- These results are from test environments—not always real-world conditions.
- Longer-term planning and error handling are still hard for AI.
What This Means for Strategy¶
- Plan Ahead: AI systems may soon be capable of completing longer tasks with little oversight.
- Pilot Projects: Start testing where AI might assist or automate longer workflows.
- Talent Planning: Expect changes in the types of roles that will benefit from human–AI collaboration.
- Risk Management: Use these benchmarks to guide safe and responsible AI adoption.
Five Year Projection¶
Starting from late 2025 (~5 hours), if the 7-month doubling rate continues:
| Date | Projected Task Horizon |
|---|---|
| November 2025 | 5 hours |
| June 2026 | 10 hours |
| January 2027 | 20 hours |
| August 2027 | 40 hours (1 week) |
| March 2028 | 80 hours (2 weeks) |
| October 2028 | 160 hours (1 month) |
| May 2029 | 320 hours (2 months) |
| December 2029 | 640 hours (4 months) |
| July 2030 | 1280 hours (8 months) |
References¶
Here are the original source references from the Metr site:
Self-Assessment Quiz¶
Test your understanding of AI task completion doubling rates.
Question 1: According to METR research, approximately how often does the length of tasks AI can complete autonomously double?
- Every 2 years
- Every 12 months
- Every 7 months
- Every 3 months
Answer
C) Every 7 months - The research shows that the length of tasks AI models can complete with 50% reliability has been doubling approximately every seven months for the past several years.
Question 2: What does "task horizon" mean in the context of AI benchmarking?
- How far in the future AI can make predictions
- The length of tasks (measured in human time) that AI can complete autonomously
- The maximum number of tasks AI can handle simultaneously
- The geographic regions where AI is deployed
Answer
B) The length of tasks (measured in human time) that AI can complete autonomously - Task horizon measures how long a task takes a skilled human professional, and then tests whether AI can complete equivalent tasks with a certain reliability threshold.
Question 3: If the 7-month doubling rate continues, what task duration might AI handle by late 2027?
- 10 minutes
- 1 hour
- About 40 hours (1 week of human work)
- 1 year of human work
Answer
C) About 40 hours (1 week of human work) - Based on the exponential projection starting from approximately 5 hours in late 2025, doubling every 7 months would reach approximately 40 hours by August 2027.
Question 4: What does a 50% success rate threshold indicate in METR's measurements?
- AI fails half the time at any task
- AI has a 50% chance of successfully completing tasks up to the measured horizon length
- Half of all AI models meet the benchmark
- AI uses 50% of available computing resources
Answer
B) AI has a 50% chance of successfully completing tasks up to the measured horizon length - The 50% success rate is the probability threshold used to define the task horizon, meaning tasks at that duration have a coin-flip chance of successful completion.
Question 5: What is an important caveat when interpreting these exponential AI capability trends?
- The measurements are completely inaccurate
- AI may still need supervision, and these are test environments not real-world conditions
- The doubling rate will definitely continue forever
- Only OpenAI models are measured
Answer
B) AI may still need supervision, and these are test environments not real-world conditions - While the trend data is valuable for planning, a 50% success rate still means significant failures, and laboratory benchmark performance may not fully translate to real-world deployment scenarios.