· Jul 14, 2025
METR reports time-horizon trends across AI benchmarks
Markdown · JSON · Original source
Published July 14, 2025, compares how long can complete tasks across different domains. METR defines a 50% as the length of tasks, measured by how long human professionals take, that an agent completes autonomously with a 50% probability. Its estimates put software, math-contest and scientific-question-answering benchmarks in a similar range, with reported doubling times of 2–6 months. METR says computer-use benchmarks have much shorter horizons, while self-driving has improved more slowly. It argues that the software results are not an outlier, but stresses that many estimates are noisy and some extend beyond the task lengths represented in their benchmarks. The post concludes that the evidence points to exponential or slightly super-exponential growth in the domains examined, while emphasizing that the benchmarks cover a limited range of domains and AI-strong tasks. METR says the evidence does not establish how these trends translate to real-world work, or whether domains with shorter horizons will catch up or stagnate. It identifies research on other domains and harder-to-automate skills as important for understanding potential bottlenecks to AI R&D automation and dangerous capabilities.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
· Article
Source details
Original source date: Jul 14, 2025 (2025-07-14T07:00:00.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-13T04:28:23.239Z. First observed: 2026-10-18T03:27:31.440Z. Retrieved: 2026-10-23T23:25:38.190Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
Text as retrieved on October 23, 2026; originally published July 14, 2025. says its evidence comes from in a limited range of domains, especially those where AIs are relatively good, and that estimates are noisy. METR says it has few domains with short and does not know whether such domains will see catch-up growth or stagnation. METR says many time-horizon estimates are extrapolations beyond the task lengths in the benchmarks; such estimates may not correspond to actual task-length ranges tested.
Open question
Will trends in domains with short horizons, or in economically important work not represented by these , resemble the trends reports?
Source-stated
METR defines a 50% time horizon by task length, measured using human professionals’ completion time, and an agent’s probability of autonomous completion.
METR defined an AI model’s 50% time horizon as the length of tasks (measured by how long they take human professionals) that it can complete autonomously with 50% probability.
Source-stated
METR reports that software, math-contest and scientific-question-answering benchmarks have similar time horizons and doubling times.
Coding, math contest, and QA benchmarks all show 30-200 minute+ horizons doubling every ~2-6 months
Source-stated
METR says agentic computer-use horizons are much shorter than coding or math horizons, and self-driving has improved more slowly.
Agentic GUI tasks (OSWorld) are still ~100 × shorter than coding or math; Tesla self-driving has improved much slower than other domains.
Source-stated
METR says the examined domains show exponential or super-exponential growth trends, while noting that its estimates are noisy and cover a limited range of domains.
We find exponential or super-exponential growth trends everywhere we look, although there is a large spread in both time horizons and growth rates.
Source-stated
METR says future work on other domains with shorter horizons may help identify bottlenecks to dangerous capabilities and AI R&D automation.
Future work on other domains with shorter horizons may provide direct evidence about the bottlenecks to dangerous capabilities and AI R&D automation.
Source and contributor context
METR · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Time horizon
METR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
AI agent
An AI system that chooses and carries out actions toward a goal, often using software tools.
Agentic AI
AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.
METR
METR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Editorial provenance
Version 1. Publication ID: fe5f97e7-8a29-d093-b1e0-11e3c8a5f572.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-13T04:23:40.226Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model.
Observation: 5977fab96c52ff51167dad747e76084bfd6799f6f0bdb3d7c380416bc6f3d884. Source text hash: 0c284fde06c17e634605af031b640bdba6ff30969b4a0664f40928d00f37215b.
Revision and source-check history
Version 1 · 2026-11-13T04:28:23.239Z
METR reports time-horizon trends across AI benchmarks
Published July 14, 2025, compares how long can complete tasks across different domains. METR defines a 50% as the length of tasks, measured by how long human professionals take, that an agent completes autonomously with a 50% probability. Its estimates put software, math-contest and scientific-question-answering benchmarks in a similar range, with reported doubling times of 2–6 months. METR says computer-use benchmarks have much shorter horizons, while self-driving has improved more slowly. It argues that the software results are not an outlier, but stresses that many estimates are noisy and some extend beyond the task lengths represented in their benchmarks. The post concludes that the evidence points to exponential or slightly super-exponential growth in the domains examined, while emphasizing that the benchmarks cover a limited range of domains and AI-strong tasks. METR says the evidence does not establish how these trends translate to real-world work, or whether domains with shorter horizons will catch up or stagnate. It identifies research on other domains and harder-to-automate skills as important for understanding potential bottlenecks to AI R&D automation and dangerous capabilities.
Original source
· Article
Source details
Original source date: Jul 14, 2025 (2025-07-14T07:00:00.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-13T04:28:23.239Z. First observed: 2026-10-18T03:27:31.440Z. Retrieved: 2026-10-23T23:25:38.190Z.
Limits
Text as retrieved on October 23, 2026; originally published July 14, 2025. says its evidence comes from in a limited range of domains, especially those where AIs are relatively good, and that estimates are noisy. METR says it has few domains with short and does not know whether such domains will see catch-up growth or stagnation. METR says many time-horizon estimates are extrapolations beyond the task lengths in the benchmarks; such estimates may not correspond to actual task-length ranges tested.
Open question
Will trends in domains with short horizons, or in economically important work not represented by these , resemble the trends reports?
Source-stated
METR defines a 50% time horizon by task length, measured using human professionals’ completion time, and an agent’s probability of autonomous completion.
METR defined an AI model’s 50% time horizon as the length of tasks (measured by how long they take human professionals) that it can complete autonomously with 50% probability.
Source-stated
METR reports that software, math-contest and scientific-question-answering benchmarks have similar time horizons and doubling times.
Coding, math contest, and QA benchmarks all show 30-200 minute+ horizons doubling every ~2-6 months
Source-stated
METR says agentic computer-use horizons are much shorter than coding or math horizons, and self-driving has improved more slowly.
Agentic GUI tasks (OSWorld) are still ~100 × shorter than coding or math; Tesla self-driving has improved much slower than other domains.
Source-stated
METR says the examined domains show exponential or super-exponential growth trends, while noting that its estimates are noisy and cover a limited range of domains.
We find exponential or super-exponential growth trends everywhere we look, although there is a large spread in both time horizons and growth rates.
Source-stated
METR says future work on other domains with shorter horizons may help identify bottlenecks to dangerous capabilities and AI R&D automation.
Future work on other domains with shorter horizons may provide direct evidence about the bottlenecks to dangerous capabilities and AI R&D automation.
Source and contributor context
METR · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Time horizon
METR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
AI agent
An AI system that chooses and carries out actions toward a goal, often using software tools.
Agentic AI
AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.
METR
METR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Editorial provenance
Version 1. Publication ID: fe5f97e7-8a29-d093-b1e0-11e3c8a5f572.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-13T04:23:40.226Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model.
Observation: 5977fab96c52ff51167dad747e76084bfd6799f6f0bdb3d7c380416bc6f3d884. Source text hash: 0c284fde06c17e634605af031b640bdba6ff30969b4a0664f40928d00f37215b.
passed · 2026-11-13T04:23:40.226Z · anthropic/claude-opus-5.5
Publication: fe5f97e7-8a29-d093-b1e0-11e3c8a5f572
attribution — pass: This is a single METR blog post with its own analysis. The reading attributes its definitions, estimates and conclusions to METR throughout and does not present them as established fact.
We find exponential or super-exponential growth trends everywhere we look
summary — pass: The reading leads with the actor and the development, and defines the 50% time horizon in plain language. It covers the domain cluster, shorter computer-use horizons, slower self-driving progress, noise and extrapolation caveats, and the limited-domain gap. Time horizon, benchmark and agent terms are defined beside the reading. The question follows from the source's stated unknowns. Omitting the video and five-year speculation does not distort the source.
we don’t know whether such domains will see catch-up growth or stagnation
date — pass: The reading gives July 14, 2025, which matches the UTC day of originalPublished and the URL path. There is no hindsight framing, and forecasts are not presented as outcomes. The evidence is a publisher page, not an archive copy, so no archive disclosure is needed.
support — pass: Each claim matches a verbatim source passage. The 2–6 month doubling times, much shorter GUI horizons, slower self-driving progress, the exponential or slightly super-exponential trend, noisy estimates, extrapolations, the limited domains and the bottleneck framing are all source-stated. The reading invents no benefits or costs.
Many of the time horizons are also extrapolations beyond the actual range of task lengths found in benchmarks
media — pass: No media was selected and no verified candidates were supplied, so there is no identity or relevance problem.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
METR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRMETR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRAn AI system that chooses and carries out actions toward a goal, often using software tools.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceMETR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
Definition referenceAI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.
Definition referenceMETR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRMETR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceMETR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
Definition referenceMETR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceMETR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRMETR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRAn AI system that chooses and carries out actions toward a goal, often using software tools.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceMETR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
Definition referenceAI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.
Definition referenceMETR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRMETR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceMETR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
Definition referenceMETR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceMETR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METR