· Jan 22, 2026
METR clarifies what its time horizon measure can and cannot say
Markdown · JSON · Original source
Originally published on January 22, 2026, this note from an unnamed researcher argues that is an imprecise measure of performance on particular tasks, not a measure of how long an AI can work independently. The author defines it as the amount of serial human labor an AI can replace at a 50% success rate, and stresses that estimates have wide error bars and vary sharply across task domains. The note also warns that tasks may not represent real-world work, and that a 50% success rate does not make a task suitable for delegation when reliability matters. The author nevertheless stands by the paper’s central trend: a long-run doubling in time horizon every 6–7 months, and a linear extrapolation predicting a one-month time horizon in 2030. The argument is that this trend matters more than the exact estimate for any particular model. But the author notes that the dataset is becoming a problem as models improve, and that longer tasks often lack human baselines; colleague Tom Cunningham has raised the possibility that misjudging their lengths could affect estimates of acceleration. The note does not establish how benchmark results translate into real-world automation or research speedup.
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: Jan 22, 2026 (2026-01-22T08:00:00.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-09T00:58:24.154Z. First observed: 2026-10-18T03:13:20.202Z. Retrieved: 2026-10-23T21:39:40.611Z.
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 January 22, 2026. The note is described as an unpolished research update and does not identify its author; it says such notes do not necessarily reflect ’s views as a whole. The author says task distributions may differ from real-world work, and that the extrapolation of trends depends on a task distribution that is not well defined.
Open question
How much would better human baselines for longer tasks change the estimated pace of AI growth?
Source-stated
The author defines time horizon as the amount of serial human labor an AI can replace with a 50% success rate, not the length of time it can work independently.
it’s the amount of serial human labor they can replace with a 50% success rate
Source-stated
The author says time-horizon estimates are imprecise, with error bars that have historically been roughly a factor of two in each direction.
Error bars have historically been a factor of ~2 in each direction
Source-stated
The author says time horizons differ substantially across domains; a follow-up found them 40–100 times lower for visual computer use tasks than for math.
time horizons are fairly similar for math, but 40-100x lower for visual computer use tasks
Source-stated
The author argues that the paper’s key result is a long-run trend of one doubling every 6–7 months, rather than an exact estimate for any particular model.
the slope of the long-run trend (one doubling every 6-7 months)
Source-stated
The author says a linear extrapolation of the trend predicted when AIs would reach a one-month time horizon: 2030.
a linear extrapolation of this trend predicting when AIs would reach 1 month / 167 working hours time horizon (2030)
Source-stated
The author reports that longer tasks often lack human baselines, and says this could affect estimates of acceleration in 2025 if their lengths are systematically misestimated.
most of the longer tasks don’t have baselines, so if we’re systematically over/under-estimating the length of long tasks we could be misjudging the degree of acceleration in 2025.
Source-stated
The author says a 50% time horizon does not by itself mean tasks below that duration can be delegated to AI.
A 50% time horizon of X hours does not mean we can delegate tasks under X hours to AIs.
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.
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: 0cf68e1c-a63f-31da-1587-e6b3bcc63c8f.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-09T00:55:53.408Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model.
Observation: 3f7637858b9dc4e09517f415e58faffba936343a9b629f774e1753b08ae7e16a. Source text hash: c7b47f4f28f50941e2e129fc8e98283a00a866a6f95ce40dab4362703a0c4618.
Revision and source-check history
Version 1 · 2026-11-09T00:58:24.154Z
METR clarifies what its time horizon measure can and cannot say
Originally published on January 22, 2026, this note from an unnamed researcher argues that is an imprecise measure of performance on particular tasks, not a measure of how long an AI can work independently. The author defines it as the amount of serial human labor an AI can replace at a 50% success rate, and stresses that estimates have wide error bars and vary sharply across task domains. The note also warns that tasks may not represent real-world work, and that a 50% success rate does not make a task suitable for delegation when reliability matters. The author nevertheless stands by the paper’s central trend: a long-run doubling in time horizon every 6–7 months, and a linear extrapolation predicting a one-month time horizon in 2030. The argument is that this trend matters more than the exact estimate for any particular model. But the author notes that the dataset is becoming a problem as models improve, and that longer tasks often lack human baselines; colleague Tom Cunningham has raised the possibility that misjudging their lengths could affect estimates of acceleration. The note does not establish how benchmark results translate into real-world automation or research speedup.
Original source
· Article
Source details
Original source date: Jan 22, 2026 (2026-01-22T08:00:00.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-09T00:58:24.154Z. First observed: 2026-10-18T03:13:20.202Z. Retrieved: 2026-10-23T21:39:40.611Z.
Limits
Text as retrieved on October 23, 2026; originally published January 22, 2026. The note is described as an unpolished research update and does not identify its author; it says such notes do not necessarily reflect ’s views as a whole. The author says task distributions may differ from real-world work, and that the extrapolation of trends depends on a task distribution that is not well defined.
Open question
How much would better human baselines for longer tasks change the estimated pace of AI growth?
Source-stated
The author defines time horizon as the amount of serial human labor an AI can replace with a 50% success rate, not the length of time it can work independently.
it’s the amount of serial human labor they can replace with a 50% success rate
Source-stated
The author says time-horizon estimates are imprecise, with error bars that have historically been roughly a factor of two in each direction.
Error bars have historically been a factor of ~2 in each direction
Source-stated
The author says time horizons differ substantially across domains; a follow-up found them 40–100 times lower for visual computer use tasks than for math.
time horizons are fairly similar for math, but 40-100x lower for visual computer use tasks
Source-stated
The author argues that the paper’s key result is a long-run trend of one doubling every 6–7 months, rather than an exact estimate for any particular model.
the slope of the long-run trend (one doubling every 6-7 months)
Source-stated
The author says a linear extrapolation of the trend predicted when AIs would reach a one-month time horizon: 2030.
a linear extrapolation of this trend predicting when AIs would reach 1 month / 167 working hours time horizon (2030)
Source-stated
The author reports that longer tasks often lack human baselines, and says this could affect estimates of acceleration in 2025 if their lengths are systematically misestimated.
most of the longer tasks don’t have baselines, so if we’re systematically over/under-estimating the length of long tasks we could be misjudging the degree of acceleration in 2025.
Source-stated
The author says a 50% time horizon does not by itself mean tasks below that duration can be delegated to AI.
A 50% time horizon of X hours does not mean we can delegate tasks under X hours to AIs.
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.
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: 0cf68e1c-a63f-31da-1587-e6b3bcc63c8f.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-09T00:55:53.408Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model.
Observation: 3f7637858b9dc4e09517f415e58faffba936343a9b629f774e1753b08ae7e16a. Source text hash: c7b47f4f28f50941e2e129fc8e98283a00a866a6f95ce40dab4362703a0c4618.
uncertain · 2026-11-09T00:55:53.408Z · anthropic/claude-opus-5.5
Publication: 0cf68e1c-a63f-31da-1587-e6b3bcc63c8f
attribution — pass: The reading attributes the note to an unnamed METR researcher. This matches the source, whose author identifies only as one of the main authors of the time horizon paper. The reading also carries the disclaimer that notes do not necessarily reflect METR's views, and it keeps assertions attributed to the author.
As one of the main authors, I often see various misinterpretations of our work.
summary — pass: The summary faithfully conveys the author's position. Time horizon is serial human labor replaced at 50% success, with wide error bars and domain variation. The author warns about delegation and benchmark representativeness, still stands by the 6–7 month doubling trend and the 2030 extrapolation, and flags Tom Cunningham's baseline concern. Time horizon and benchmark are defined beside the reading. The question follows directly from the source's stated concern.
I still believe in the core results, I believe that many people to some extent both overstate the precision of our time horizon measurements
date — pass: The reading states January 22, 2026. This matches the UTC calendar day of originalPublished (2026-01-22T08:00Z) and the URL path. The 2030 one-month horizon is presented as a linear-extrapolation prediction, not an outcome. There is no hindsight framing, and the reading does not rely on later material.
support — uncertain: Most claims match their excerpts. One does not: the third claim says visual computer use time horizons were 40–100 times lower than for math. The source says math was fairly similar to the original software/research tasks, while visual computer use was 40–100x lower, which is a comparison against the original task domain rather than against math. Smallest repair: change the comparison to the original software and research tasks, noting that math was similar.
time horizons are fairly similar for math, but 40-100x lower for visual computer use tasks
media — pass: No media was selected and no verified media candidates were supplied, so there is nothing to mismatch.
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 METRMETR'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 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 referenceMETR studies what AI systems can do and how reliably they complete tasks, including work with human developers.
Explore METRMETR'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 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 reference