{"id":"15c4c9412539b52e849b7b8e40c3d2a64ff67d1b587d21cc97b7ef8e443656d4","title":"METR reports revised estimates of AI models’ time horizons","argument":"METR published Time Horizon 1.1 on January 29, 2026, revising its estimates of how long AI models can carry out tasks. The measure, called “time horizon,” tracks task length using how long people take to complete the task. METR updated its task suite and moved its evaluation infrastructure from Vivaria to Inspect. It says the revised estimates generally fall within the earlier confidence intervals, but the growth trend since 2023 is faster under the new method: the estimated doubling time is 131 days, compared with 165 days under the earlier one. METR attributes the changed trend mainly to revised estimates for older and newer models, and says the task suite’s composition affects the trend. It considers the new suite an improvement, while acknowledging it estimates a slightly different underlying trend. The expanded suite gives tighter estimates, especially for newer models, but METR says the confidence intervals remain wide and that it is prioritizing evaluations able to measure very strong models. The reported estimates also rely on estimated human completion times for most of the suite’s longest tasks. How will METR’s continued changes to its task selection affect comparisons of time-horizon growth over time?","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"METR says the confidence intervals remain very wide, and that it is working to add more long tasks. Text as retrieved on October 23, 2026; originally published January 29, 2026.","speaker":{"name":"METR","roleAtSource":"Publisher"},"source":{"title":"Time Horizon 1.1","publisher":"METR","format":"Article","url":"https://metr.org/blog/2026-01-29-time-horizon-1-1/","publishedAt":"2026-01-29T08:00:00.000Z","dateBasis":"Date supplied by the publisher’s feed."},"evidence":[{"text":"METR says the revised time-horizon estimates generally fall within the confidence intervals from its earlier estimates.","excerpt":"The new estimates generally lie within the confidence intervals from the TH1 time horizons.","url":"https://metr.org/blog/2026-01-29-time-horizon-1-1/","relationship":"Source-stated"},{"text":"METR reports that the estimated post-2023 doubling time is 131 days under TH1.1 and 165 days under TH1.","excerpt":"The post-2023 doubling-time is 131 days under TH1.1, compared to 165 days under TH1","url":"https://metr.org/blog/2026-01-29-time-horizon-1-1/","relationship":"Source-stated"},{"text":"METR says the changed trend reflects lower estimates for older models and higher estimates for recent models, and that task composition affects the trend.","excerpt":"The change in trend is due to our estimates of older models shifting down","url":"https://metr.org/blog/2026-01-29-time-horizon-1-1/","relationship":"Source-stated"},{"text":"METR says its new task suite is an improvement, but estimates a slightly different underlying trend; it also says the confidence intervals remain wide.","excerpt":"We believe the new task suite to be overall an improvement, and so worth this slight change in the quantity being estimated.","url":"https://metr.org/blog/2026-01-29-time-horizon-1-1/","relationship":"Source-stated"},{"text":"METR reports that it measured human baseline times for only 5 of its 31 long tasks; the rest use estimated times.","excerpt":"It is also notable that we measured human baseline times for only 5 of our 31 long (8h+) tasks. The remainder use estimated times.","url":"https://metr.org/blog/2026-01-29-time-horizon-1-1/","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://metr.org/blog/2026-01-29-time-horizon-1-1/","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How will METR’s continued changes to its task selection affect comparisons of time-horizon growth over time?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"METR reports revised estimates of AI models’ time horizons","summary":"METR published Time Horizon 1.1 on January 29, 2026, revising its estimates of how long AI models can carry out tasks. The measure, called “time horizon,” tracks task length using how long people take to complete the task. METR updated its task suite and moved its evaluation infrastructure from Vivaria to Inspect. It says the revised estimates generally fall within the earlier confidence intervals, but the growth trend since 2023 is faster under the new method: the estimated doubling time is 131 days, compared with 165 days under the earlier one. METR attributes the changed trend mainly to revised estimates for older and newer models, and says the task suite’s composition affects the trend. It considers the new suite an improvement, while acknowledging it estimates a slightly different underlying trend. The expanded suite gives tighter estimates, especially for newer models, but METR says the confidence intervals remain wide and that it is prioritizing evaluations able to measure very strong models. The reported estimates also rely on estimated human completion times for most of the suite’s longest tasks. How will METR’s continued changes to its task selection affect comparisons of time-horizon growth over time?","question":"How will METR’s continued changes to its task selection affect comparisons of time-horizon growth over time?","claims":[{"text":"METR says the revised time-horizon estimates generally fall within the confidence intervals from its earlier estimates.","excerpt":"The new estimates generally lie within the confidence intervals from the TH1 time horizons.","support":"source-stated"},{"text":"METR reports that the estimated post-2023 doubling time is 131 days under TH1.1 and 165 days under TH1.","excerpt":"The post-2023 doubling-time is 131 days under TH1.1, compared to 165 days under TH1","support":"source-stated"},{"text":"METR says the changed trend reflects lower estimates for older models and higher estimates for recent models, and that task composition affects the trend.","excerpt":"The change in trend is due to our estimates of older models shifting down","support":"source-stated"},{"text":"METR says its new task suite is an improvement, but estimates a slightly different underlying trend; it also says the confidence intervals remain wide.","excerpt":"We believe the new task suite to be overall an improvement, and so worth this slight change in the quantity being estimated.","support":"source-stated"},{"text":"METR reports that it measured human baseline times for only 5 of its 31 long tasks; the rest use estimated times.","excerpt":"It is also notable that we measured human baseline times for only 5 of our 31 long (8h+) tasks. The remainder use estimated times.","support":"source-stated"}],"limitations":["METR says the confidence intervals remain very wide, and that it is working to add more long tasks.","Text as retrieved on October 23, 2026; originally published January 29, 2026."],"mediaId":null,"correctionReason":null,"id":"15c4c9412539b52e849b7b8e40c3d2a64ff67d1b587d21cc97b7ef8e443656d4","publicationId":"dc45c7df-5bf6-a535-9730-85d63837fe07","version":1,"sourceUrl":"https://metr.org/blog/2026-01-29-time-horizon-1-1/","sourceTitle":"Time Horizon 1.1","publisher":"METR","publishedAt":"2026-01-29T08:00:00.000Z","originalPublished":{"value":"2026-01-29T08:00:00.000Z","precision":"instant","basis":"backfill:feed_published"},"firstObservedAt":"2026-10-18T03:11:20.936Z","createdAt":"2026-11-08T10:48:23.725Z","historical":true,"entityIds":["metr"],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://metr.org/blog/2026-01-29-time-horizon-1-1/","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"9999669e7f4268d604a31b64589ee1440d85501eec5fb3c1b9b60c44492409c5","rubricVersion":"unrecorded","provider":"vercel","model":"typesafe-ai/jev","reasonCodes":["bridge_admitted_by_model"],"confidence":0.8,"decidedAt":"2026-10-23T23:37:42.659Z"},"provenance":{"textHash":"1fe13d89bc9c58dd1c2bf8cc6cfeb52cf31c1705702dcbbc6a2fc7d70559b175","observationId":"118ad060da4fb71c489f4bc057f330d2d3dfd9290dbf5144f933c250a3b69c51","retrievedAt":"2026-10-23T23:39:36.528Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"15012eea-ea7c-bd24-12b3-81ea5c1ef28b"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"588d87b2-28d4-00e6-9df2-8e8c267b110c"}},"audit":{"status":"passed","reviewedAt":"2026-11-08T10:43:38.883Z","model":"anthropic/claude-opus-5.5"}},"canonicalUrl":"https://escapemeridian.com/readings/metr-reports-revised-estimates-of-ai-models-time-horizons-15c4c941","formats":{"html":"/readings/metr-reports-revised-estimates-of-ai-models-time-horizons-15c4c941","markdown":"/readings/metr-reports-revised-estimates-of-ai-models-time-horizons-15c4c941.md","json":"/readings/metr-reports-revised-estimates-of-ai-models-time-horizons-15c4c941.json"},"attribution":"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.","maintainedGuideLinks":[],"readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[{"id":"term-time-horizon","name":"Time horizon","explanation":"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.","referenceUrl":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/"},{"id":"term-confidence-interval","name":"Confidence interval","explanation":"A range of values consistent with the data at a stated level of confidence.","referenceUrl":"https://en.wikipedia.org/wiki/Confidence_interval"}],"entities":[{"id":"metr","name":"METR","kind":"organization","context":"METR studies what AI systems can do and how reliably they complete tasks, including work with human developers.","referenceUrls":["https://metr.org/"],"path":"/entities/metr","json":"/entities/metr.json","markdown":"/entities/metr.md"},{"id":"time","name":"TIME","kind":"organization","context":"TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.","referenceUrls":["https://time.com/"],"path":"/entities/time","json":"/entities/time.json","markdown":"/entities/time.md"}]},"researchQuestions":[{"id":"agi","title":"Is AGI a few years away?","question":"AGI within a few years: progress, limits, timelines or forecasts for broadly capable systems, including disagreement.","url":"/questions/agi","edition":"product-management-read-2026-11-04"},{"id":"ai-delegation-reliability","title":"When can work be delegated without checking every step?","question":"When can work be delegated without checking every step?","url":"/questions/ai-delegation-reliability","edition":"product-management-read-2026-11-04"}],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"15c4c9412539b52e849b7b8e40c3d2a64ff67d1b587d21cc97b7ef8e443656d4","revision":"dc45c7df-5bf6-a535-9730-85d63837fe07","revisionKind":"publication-id","title":"METR reports revised estimates of AI models’ time horizons","url":"https://escapemeridian.com/readings/metr-reports-revised-estimates-of-ai-models-time-horizons-15c4c941"},"source":{"type":"source","id":"https://metr.org/blog/2026-01-29-time-horizon-1-1/","revision":"1fe13d89bc9c58dd1c2bf8cc6cfeb52cf31c1705702dcbbc6a2fc7d70559b175","revisionKind":"sha256-source-text","title":"Time Horizon 1.1","url":"https://metr.org/blog/2026-01-29-time-horizon-1-1/"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter."},"revisions":[{"id":"15c4c9412539b52e849b7b8e40c3d2a64ff67d1b587d21cc97b7ef8e443656d4","title":"METR reports revised estimates of AI models’ time horizons","argument":"METR published Time Horizon 1.1 on January 29, 2026, revising its estimates of how long AI models can carry out tasks. The measure, called “time horizon,” tracks task length using how long people take to complete the task. METR updated its task suite and moved its evaluation infrastructure from Vivaria to Inspect. It says the revised estimates generally fall within the earlier confidence intervals, but the growth trend since 2023 is faster under the new method: the estimated doubling time is 131 days, compared with 165 days under the earlier one. METR attributes the changed trend mainly to revised estimates for older and newer models, and says the task suite’s composition affects the trend. It considers the new suite an improvement, while acknowledging it estimates a slightly different underlying trend. The expanded suite gives tighter estimates, especially for newer models, but METR says the confidence intervals remain wide and that it is prioritizing evaluations able to measure very strong models. The reported estimates also rely on estimated human completion times for most of the suite’s longest tasks. How will METR’s continued changes to its task selection affect comparisons of time-horizon growth over time?","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"METR says the confidence intervals remain very wide, and that it is working to add more long tasks. Text as retrieved on October 23, 2026; originally published January 29, 2026.","speaker":{"name":"METR","roleAtSource":"Publisher"},"source":{"title":"Time Horizon 1.1","publisher":"METR","format":"Article","url":"https://metr.org/blog/2026-01-29-time-horizon-1-1/","publishedAt":"2026-01-29T08:00:00.000Z","dateBasis":"Date supplied by the publisher’s feed."},"evidence":[{"text":"METR says the revised time-horizon estimates generally fall within the confidence intervals from its earlier estimates.","excerpt":"The new estimates generally lie within the confidence intervals from the TH1 time horizons.","url":"https://metr.org/blog/2026-01-29-time-horizon-1-1/","relationship":"Source-stated"},{"text":"METR reports that the estimated post-2023 doubling time is 131 days under TH1.1 and 165 days under TH1.","excerpt":"The post-2023 doubling-time is 131 days under TH1.1, compared to 165 days under TH1","url":"https://metr.org/blog/2026-01-29-time-horizon-1-1/","relationship":"Source-stated"},{"text":"METR says the changed trend reflects lower estimates for older models and higher estimates for recent models, and that task composition affects the trend.","excerpt":"The change in trend is due to our estimates of older models shifting down","url":"https://metr.org/blog/2026-01-29-time-horizon-1-1/","relationship":"Source-stated"},{"text":"METR says its new task suite is an improvement, but estimates a slightly different underlying trend; it also says the confidence intervals remain wide.","excerpt":"We believe the new task suite to be overall an improvement, and so worth this slight change in the quantity being estimated.","url":"https://metr.org/blog/2026-01-29-time-horizon-1-1/","relationship":"Source-stated"},{"text":"METR reports that it measured human baseline times for only 5 of its 31 long tasks; the rest use estimated times.","excerpt":"It is also notable that we measured human baseline times for only 5 of our 31 long (8h+) tasks. The remainder use estimated times.","url":"https://metr.org/blog/2026-01-29-time-horizon-1-1/","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://metr.org/blog/2026-01-29-time-horizon-1-1/","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How will METR’s continued changes to its task selection affect comparisons of time-horizon growth over time?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"METR reports revised estimates of AI models’ time horizons","summary":"METR published Time Horizon 1.1 on January 29, 2026, revising its estimates of how long AI models can carry out tasks. The measure, called “time horizon,” tracks task length using how long people take to complete the task. METR updated its task suite and moved its evaluation infrastructure from Vivaria to Inspect. It says the revised estimates generally fall within the earlier confidence intervals, but the growth trend since 2023 is faster under the new method: the estimated doubling time is 131 days, compared with 165 days under the earlier one. METR attributes the changed trend mainly to revised estimates for older and newer models, and says the task suite’s composition affects the trend. It considers the new suite an improvement, while acknowledging it estimates a slightly different underlying trend. The expanded suite gives tighter estimates, especially for newer models, but METR says the confidence intervals remain wide and that it is prioritizing evaluations able to measure very strong models. The reported estimates also rely on estimated human completion times for most of the suite’s longest tasks. How will METR’s continued changes to its task selection affect comparisons of time-horizon growth over time?","question":"How will METR’s continued changes to its task selection affect comparisons of time-horizon growth over time?","claims":[{"text":"METR says the revised time-horizon estimates generally fall within the confidence intervals from its earlier estimates.","excerpt":"The new estimates generally lie within the confidence intervals from the TH1 time horizons.","support":"source-stated"},{"text":"METR reports that the estimated post-2023 doubling time is 131 days under TH1.1 and 165 days under TH1.","excerpt":"The post-2023 doubling-time is 131 days under TH1.1, compared to 165 days under TH1","support":"source-stated"},{"text":"METR says the changed trend reflects lower estimates for older models and higher estimates for recent models, and that task composition affects the trend.","excerpt":"The change in trend is due to our estimates of older models shifting down","support":"source-stated"},{"text":"METR says its new task suite is an improvement, but estimates a slightly different underlying trend; it also says the confidence intervals remain wide.","excerpt":"We believe the new task suite to be overall an improvement, and so worth this slight change in the quantity being estimated.","support":"source-stated"},{"text":"METR reports that it measured human baseline times for only 5 of its 31 long tasks; the rest use estimated times.","excerpt":"It is also notable that we measured human baseline times for only 5 of our 31 long (8h+) tasks. The remainder use estimated times.","support":"source-stated"}],"limitations":["METR says the confidence intervals remain very wide, and that it is working to add more long tasks.","Text as retrieved on October 23, 2026; originally published January 29, 2026."],"mediaId":null,"correctionReason":null,"id":"15c4c9412539b52e849b7b8e40c3d2a64ff67d1b587d21cc97b7ef8e443656d4","publicationId":"dc45c7df-5bf6-a535-9730-85d63837fe07","version":1,"sourceUrl":"https://metr.org/blog/2026-01-29-time-horizon-1-1/","sourceTitle":"Time Horizon 1.1","publisher":"METR","publishedAt":"2026-01-29T08:00:00.000Z","originalPublished":{"value":"2026-01-29T08:00:00.000Z","precision":"instant","basis":"backfill:feed_published"},"firstObservedAt":"2026-10-18T03:11:20.936Z","createdAt":"2026-11-08T10:48:23.725Z","historical":true,"entityIds":["metr"],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://metr.org/blog/2026-01-29-time-horizon-1-1/","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"9999669e7f4268d604a31b64589ee1440d85501eec5fb3c1b9b60c44492409c5","rubricVersion":"unrecorded","provider":"vercel","model":"typesafe-ai/jev","reasonCodes":["bridge_admitted_by_model"],"confidence":0.8,"decidedAt":"2026-10-23T23:37:42.659Z"},"provenance":{"textHash":"1fe13d89bc9c58dd1c2bf8cc6cfeb52cf31c1705702dcbbc6a2fc7d70559b175","observationId":"118ad060da4fb71c489f4bc057f330d2d3dfd9290dbf5144f933c250a3b69c51","retrievedAt":"2026-10-23T23:39:36.528Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"15012eea-ea7c-bd24-12b3-81ea5c1ef28b"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"588d87b2-28d4-00e6-9df2-8e8c267b110c"}},"audit":{"status":"passed","reviewedAt":"2026-11-08T10:43:38.883Z","model":"anthropic/claude-opus-5.5"}}}],"audits":[{"publicationId":"dc45c7df-5bf6-a535-9730-85d63837fe07","verdict":"passed","findings":[{"dimension":"attribution","status":"pass","explanation":"The text is a single METR blog post announcing Time Horizon 1.1. The reading attributes the revised estimates, trend figures and judgments to METR throughout.","excerpt":"We’re releasing a new version of our time horizon estimates (TH1.1), using more tasks and a new eval infrastructure."},{"dimension":"summary","status":"pass","explanation":"The summary accurately states the actor and the change: a new task suite and the move from Vivaria to Inspect. It gives the faster post-2023 trend, the causes, METR's caveats and the reliance on estimated human times for long tasks. Time horizon and confidence interval are defined on the page. The opening phrase about how long models can carry out tasks is loose, but the next sentence clarifies that task length is measured by human completion time. The question follows from the evidence.","excerpt":"The trend in time horizon is somewhat sensitive to task composition."},{"dimension":"date","status":"pass","explanation":"originalPublished is 2026-01-29T08:00Z, and the reading's January 29, 2026 is the UTC day, consistent with the URL. Nothing post-publication or hindsight framing appears. METR's future work is presented as plans. The evidence is a publisher page, not an archive copy.","excerpt":null},{"dimension":"support","status":"pass","explanation":"Each claim matches the source text: estimates within earlier intervals, doubling times of 131 vs 165 days, the causes of the trend change, the improvement and its tradeoff, wide intervals, and 5 of 31 long tasks being human-baselined. Tighter estimates for newer models are supported. No unsupported benefits or impacts are implied.","excerpt":"This gives us significantly tighter confidence intervals on the time horizon of more recent models"},{"dimension":"media","status":"pass","explanation":"No media was selected, and no verified media candidates were supplied, so there is no identity or relevance problem.","excerpt":null}],"createdAt":"2026-11-08T10:43:38.883Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}