{"id":"c291dce75d86a17df5d2beae35db6c3e3e29efc17d8c877e27f4cbcc501cb33b","title":"METR measures AI agents by the length of tasks they can complete","argument":"METR published this work on March 19, 2025. The authors propose measuring AI agents by the length of tasks they can complete, using the time those tasks take human professionals as a measure of difficulty. They report that agents’ measured task-completion horizons grew exponentially over the preceding six years, with a doubling time of around seven months.\n\nMETR’s measure estimates the task length at which a model reaches a chosen success probability; the post focuses on 50% reliability. The authors say that current models can succeed on some tasks taking experts hours, but reliably complete tasks only up to a few minutes long. They argue that this measure helps explain why strong benchmark results do not necessarily translate into robust help with day-to-day work.\n\nThe authors forecast that, if the past trend continues, agents could perform a wide range of week-long tasks within two to four years, and month-long projects by the end of the decade. These are forecasts, not demonstrated outcomes. METR says the estimate depends on the tasks and human performance measured, and notes uncertainty about whether the trend will continue. How well does task length measured on these evaluations predict success on real-world work?","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 24, 2026; originally published March 19, 2025. The retrieved page includes later updates and measurements; this reading uses the material presented as reflecting the state of the data at original publication and excludes later additions. The authors state that task-horizon estimates depend on task selection and how human performance is measured, and that future trend changes and external validity are major sources of uncertainty.","speaker":{"name":"METR","roleAtSource":"Publisher"},"source":{"title":"Measuring AI Ability to Complete Long Software Tasks","publisher":"METR","format":"Article","url":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","publishedAt":"2025-03-19T07:00:00.000Z","dateBasis":"Date supplied by the publisher’s feed."},"evidence":[{"text":"METR proposes characterizing AI performance by the length of tasks agents can complete, measured using the time those tasks take human professionals.","excerpt":"We propose measuring AI performance in terms of the length of tasks AI agents can complete.","url":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","relationship":"Source-stated"},{"text":"METR reports an exponential increase in measured task-completion length, with an estimated doubling time of around seven months.","excerpt":"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 7 months for the last 6 years.","url":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","relationship":"Source-stated"},{"text":"METR says the task-length measure depends on methodological choices, including which tasks and human performance are measured.","excerpt":"Our estimate of the length of tasks that an agent can complete depends on methodological choices like the tasks used and the humans whose performance is measured.","url":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","relationship":"Source-stated"},{"text":"METR forecasts that, if the measured trend continues for two to four more years, generalist autonomous agents could perform a wide range of week-long tasks.","excerpt":"If the measured trend from the past 6 years continues for 2-4 more years, generalist autonomous agents will be capable of performing a wide range of week-long tasks.","url":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","relationship":"Source-stated"},{"text":"Escape Meridian reads the gap between measured task performance and day-to-day work as an unresolved question about how well these evaluations predict success outside the tested tasks.","excerpt":"it is unclear how this corresponds to real-world impact.","url":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","relationship":"Escape Meridian interpretation"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How well does task length measured on these evaluations predict success on real-world work?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"METR measures AI agents by the length of tasks they can complete","summary":"METR published this work on March 19, 2025. The authors propose measuring AI agents by the length of tasks they can complete, using the time those tasks take human professionals as a measure of difficulty. They report that agents’ measured task-completion horizons grew exponentially over the preceding six years, with a doubling time of around seven months.\n\nMETR’s measure estimates the task length at which a model reaches a chosen success probability; the post focuses on 50% reliability. The authors say that current models can succeed on some tasks taking experts hours, but reliably complete tasks only up to a few minutes long. They argue that this measure helps explain why strong benchmark results do not necessarily translate into robust help with day-to-day work.\n\nThe authors forecast that, if the past trend continues, agents could perform a wide range of week-long tasks within two to four years, and month-long projects by the end of the decade. These are forecasts, not demonstrated outcomes. METR says the estimate depends on the tasks and human performance measured, and notes uncertainty about whether the trend will continue. How well does task length measured on these evaluations predict success on real-world work?","question":"How well does task length measured on these evaluations predict success on real-world work?","claims":[{"text":"METR proposes characterizing AI performance by the length of tasks agents can complete, measured using the time those tasks take human professionals.","excerpt":"We propose measuring AI performance in terms of the length of tasks AI agents can complete.","support":"source-stated"},{"text":"METR reports an exponential increase in measured task-completion length, with an estimated doubling time of around seven months.","excerpt":"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 7 months for the last 6 years.","support":"source-stated"},{"text":"METR says the task-length measure depends on methodological choices, including which tasks and human performance are measured.","excerpt":"Our estimate of the length of tasks that an agent can complete depends on methodological choices like the tasks used and the humans whose performance is measured.","support":"source-stated"},{"text":"METR forecasts that, if the measured trend continues for two to four more years, generalist autonomous agents could perform a wide range of week-long tasks.","excerpt":"If the measured trend from the past 6 years continues for 2-4 more years, generalist autonomous agents will be capable of performing a wide range of week-long tasks.","support":"source-stated"},{"text":"Escape Meridian reads the gap between measured task performance and day-to-day work as an unresolved question about how well these evaluations predict success outside the tested tasks.","excerpt":"it is unclear how this corresponds to real-world impact.","support":"interpretation"}],"limitations":["Text as retrieved on October 24, 2026; originally published March 19, 2025.","The retrieved page includes later updates and measurements; this reading uses the material presented as reflecting the state of the data at original publication and excludes later additions.","The authors state that task-horizon estimates depend on task selection and how human performance is measured, and that future trend changes and external validity are major sources of uncertainty."],"mediaId":null,"correctionReason":null,"id":"c291dce75d86a17df5d2beae35db6c3e3e29efc17d8c877e27f4cbcc501cb33b","publicationId":"b30e1e91-b875-ed2d-0eaa-13f4ebb91660","version":1,"sourceUrl":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","sourceTitle":"Measuring AI Ability to Complete Long Software Tasks","publisher":"METR","publishedAt":"2025-03-19T07:00:00.000Z","originalPublished":{"value":"2025-03-19T07:00:00.000Z","precision":"instant","basis":"backfill:feed_published"},"firstObservedAt":"2026-10-18T03:33:21.122Z","createdAt":"2026-10-29T15:58:18.398Z","historical":true,"entityIds":["metr"],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"7e3bd15974cf94784aaaa2033a541da248394d02998871545d2e5538bb9132ff","rubricVersion":"unrecorded","provider":"vercel","model":"typesafe-ai/jev","reasonCodes":["bridge_admitted_by_model"],"confidence":0.73,"decidedAt":"2026-10-24T00:17:36.111Z"},"provenance":{"textHash":"fc96700d2a2fe6322f22f004bfa5e6c982f683e4bccc8e099557377fe0d354f4","observationId":"11596838bed68a7b18fb49e90b82edcf8cde1b655791ef1d43bad727bd04e08c","retrievedAt":"2026-10-24T00:19:35.572Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"ce1d2f58-822d-0a7a-858e-a5110c8c2580"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"f63e9995-d2d9-b977-457a-80d379cb3fc2"}},"audit":{"status":"passed","reviewedAt":"2026-10-29T15:56:45.326Z","model":"anthropic/claude-opus-5.5"}},"canonicalUrl":"https://escapemeridian.com/readings/metr-measures-ai-agents-by-the-length-of-tasks-they-can-complete-c291dce7","formats":{"html":"/readings/metr-measures-ai-agents-by-the-length-of-tasks-they-can-complete-c291dce7","markdown":"/readings/metr-measures-ai-agents-by-the-length-of-tasks-they-can-complete-c291dce7.md","json":"/readings/metr-measures-ai-agents-by-the-length-of-tasks-they-can-complete-c291dce7.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.","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-benchmark","name":"Benchmark","explanation":"A defined set of tests used to compare systems. Results depend on the tasks and measures selected.","referenceUrl":"https://arxiv.org/abs/2211.09110"}],"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"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"c291dce75d86a17df5d2beae35db6c3e3e29efc17d8c877e27f4cbcc501cb33b","revision":"b30e1e91-b875-ed2d-0eaa-13f4ebb91660","revisionKind":"publication-id","title":"METR measures AI agents by the length of tasks they can complete","url":"https://escapemeridian.com/readings/metr-measures-ai-agents-by-the-length-of-tasks-they-can-complete-c291dce7"},"source":{"type":"source","id":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","revision":"fc96700d2a2fe6322f22f004bfa5e6c982f683e4bccc8e099557377fe0d354f4","revisionKind":"sha256-source-text","title":"Measuring AI Ability to Complete Long Software Tasks","url":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/"},"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":"c291dce75d86a17df5d2beae35db6c3e3e29efc17d8c877e27f4cbcc501cb33b","title":"METR measures AI agents by the length of tasks they can complete","argument":"METR published this work on March 19, 2025. The authors propose measuring AI agents by the length of tasks they can complete, using the time those tasks take human professionals as a measure of difficulty. They report that agents’ measured task-completion horizons grew exponentially over the preceding six years, with a doubling time of around seven months.\n\nMETR’s measure estimates the task length at which a model reaches a chosen success probability; the post focuses on 50% reliability. The authors say that current models can succeed on some tasks taking experts hours, but reliably complete tasks only up to a few minutes long. They argue that this measure helps explain why strong benchmark results do not necessarily translate into robust help with day-to-day work.\n\nThe authors forecast that, if the past trend continues, agents could perform a wide range of week-long tasks within two to four years, and month-long projects by the end of the decade. These are forecasts, not demonstrated outcomes. METR says the estimate depends on the tasks and human performance measured, and notes uncertainty about whether the trend will continue. How well does task length measured on these evaluations predict success on real-world work?","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 24, 2026; originally published March 19, 2025. The retrieved page includes later updates and measurements; this reading uses the material presented as reflecting the state of the data at original publication and excludes later additions. The authors state that task-horizon estimates depend on task selection and how human performance is measured, and that future trend changes and external validity are major sources of uncertainty.","speaker":{"name":"METR","roleAtSource":"Publisher"},"source":{"title":"Measuring AI Ability to Complete Long Software Tasks","publisher":"METR","format":"Article","url":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","publishedAt":"2025-03-19T07:00:00.000Z","dateBasis":"Date supplied by the publisher’s feed."},"evidence":[{"text":"METR proposes characterizing AI performance by the length of tasks agents can complete, measured using the time those tasks take human professionals.","excerpt":"We propose measuring AI performance in terms of the length of tasks AI agents can complete.","url":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","relationship":"Source-stated"},{"text":"METR reports an exponential increase in measured task-completion length, with an estimated doubling time of around seven months.","excerpt":"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 7 months for the last 6 years.","url":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","relationship":"Source-stated"},{"text":"METR says the task-length measure depends on methodological choices, including which tasks and human performance are measured.","excerpt":"Our estimate of the length of tasks that an agent can complete depends on methodological choices like the tasks used and the humans whose performance is measured.","url":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","relationship":"Source-stated"},{"text":"METR forecasts that, if the measured trend continues for two to four more years, generalist autonomous agents could perform a wide range of week-long tasks.","excerpt":"If the measured trend from the past 6 years continues for 2-4 more years, generalist autonomous agents will be capable of performing a wide range of week-long tasks.","url":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","relationship":"Source-stated"},{"text":"Escape Meridian reads the gap between measured task performance and day-to-day work as an unresolved question about how well these evaluations predict success outside the tested tasks.","excerpt":"it is unclear how this corresponds to real-world impact.","url":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","relationship":"Escape Meridian interpretation"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How well does task length measured on these evaluations predict success on real-world work?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"METR measures AI agents by the length of tasks they can complete","summary":"METR published this work on March 19, 2025. The authors propose measuring AI agents by the length of tasks they can complete, using the time those tasks take human professionals as a measure of difficulty. They report that agents’ measured task-completion horizons grew exponentially over the preceding six years, with a doubling time of around seven months.\n\nMETR’s measure estimates the task length at which a model reaches a chosen success probability; the post focuses on 50% reliability. The authors say that current models can succeed on some tasks taking experts hours, but reliably complete tasks only up to a few minutes long. They argue that this measure helps explain why strong benchmark results do not necessarily translate into robust help with day-to-day work.\n\nThe authors forecast that, if the past trend continues, agents could perform a wide range of week-long tasks within two to four years, and month-long projects by the end of the decade. These are forecasts, not demonstrated outcomes. METR says the estimate depends on the tasks and human performance measured, and notes uncertainty about whether the trend will continue. How well does task length measured on these evaluations predict success on real-world work?","question":"How well does task length measured on these evaluations predict success on real-world work?","claims":[{"text":"METR proposes characterizing AI performance by the length of tasks agents can complete, measured using the time those tasks take human professionals.","excerpt":"We propose measuring AI performance in terms of the length of tasks AI agents can complete.","support":"source-stated"},{"text":"METR reports an exponential increase in measured task-completion length, with an estimated doubling time of around seven months.","excerpt":"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 7 months for the last 6 years.","support":"source-stated"},{"text":"METR says the task-length measure depends on methodological choices, including which tasks and human performance are measured.","excerpt":"Our estimate of the length of tasks that an agent can complete depends on methodological choices like the tasks used and the humans whose performance is measured.","support":"source-stated"},{"text":"METR forecasts that, if the measured trend continues for two to four more years, generalist autonomous agents could perform a wide range of week-long tasks.","excerpt":"If the measured trend from the past 6 years continues for 2-4 more years, generalist autonomous agents will be capable of performing a wide range of week-long tasks.","support":"source-stated"},{"text":"Escape Meridian reads the gap between measured task performance and day-to-day work as an unresolved question about how well these evaluations predict success outside the tested tasks.","excerpt":"it is unclear how this corresponds to real-world impact.","support":"interpretation"}],"limitations":["Text as retrieved on October 24, 2026; originally published March 19, 2025.","The retrieved page includes later updates and measurements; this reading uses the material presented as reflecting the state of the data at original publication and excludes later additions.","The authors state that task-horizon estimates depend on task selection and how human performance is measured, and that future trend changes and external validity are major sources of uncertainty."],"mediaId":null,"correctionReason":null,"id":"c291dce75d86a17df5d2beae35db6c3e3e29efc17d8c877e27f4cbcc501cb33b","publicationId":"b30e1e91-b875-ed2d-0eaa-13f4ebb91660","version":1,"sourceUrl":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","sourceTitle":"Measuring AI Ability to Complete Long Software Tasks","publisher":"METR","publishedAt":"2025-03-19T07:00:00.000Z","originalPublished":{"value":"2025-03-19T07:00:00.000Z","precision":"instant","basis":"backfill:feed_published"},"firstObservedAt":"2026-10-18T03:33:21.122Z","createdAt":"2026-10-29T15:58:18.398Z","historical":true,"entityIds":["metr"],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"7e3bd15974cf94784aaaa2033a541da248394d02998871545d2e5538bb9132ff","rubricVersion":"unrecorded","provider":"vercel","model":"typesafe-ai/jev","reasonCodes":["bridge_admitted_by_model"],"confidence":0.73,"decidedAt":"2026-10-24T00:17:36.111Z"},"provenance":{"textHash":"fc96700d2a2fe6322f22f004bfa5e6c982f683e4bccc8e099557377fe0d354f4","observationId":"11596838bed68a7b18fb49e90b82edcf8cde1b655791ef1d43bad727bd04e08c","retrievedAt":"2026-10-24T00:19:35.572Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"ce1d2f58-822d-0a7a-858e-a5110c8c2580"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"f63e9995-d2d9-b977-457a-80d379cb3fc2"}},"audit":{"status":"passed","reviewedAt":"2026-10-29T15:56:45.326Z","model":"anthropic/claude-opus-5.5"}}}],"audits":[{"publicationId":"b30e1e91-b875-ed2d-0eaa-13f4ebb91660","verdict":"passed","findings":[{"dimension":"attribution","status":"pass","explanation":"The proposal, results and forecasts are attributed to METR and its authors, matching the publisher page. The interpretive claim is labeled as Escape Meridian's reading.","excerpt":"We propose measuring AI performance in terms of the length of tasks AI agents can complete."},{"dimension":"summary","status":"pass","explanation":"The reading identifies METR and explains the metric in plain language: task length measured by human professional time, with a chosen success probability. It covers the seven-month doubling and the gap between hours-long successes and reliable minutes-long work. The question follows from the source's stated uncertainty about real-world impact.","excerpt":"are capable of some tasks that take even expert humans hours, but can only reliably complete tasks of up to a few minutes long"},{"dimension":"date","status":"pass","explanation":"originalPublished is 2025-03-19T07:00Z. Stated as March 19, 2025, it matches the UTC day and the URL. The page carries later update notes and TH1.1 and Nov 2025 material. The reading discloses these and excludes them, relying on claims the note says reflect the original publication. Forecasts are framed as conditional and not as outcomes. This is a publisher page, so no archive disclosure is needed.","excerpt":"The static figures and some claims in the text (e.g. the doubling time) reflect the state of the data at the time of original publication."},{"dimension":"support","status":"pass","explanation":"Each claim is supported by a verbatim source passage. The month-long-projects-by-end-of-decade forecast keeps the source's conditional framing. The trend-continuation uncertainty is supported by the sensitivity-plot note. The interpretation is labeled and grounded in the source's statement on real-world impact.","excerpt":"Note that this plot does not account for future changes in the trend or external validity concerns, which are responsible for the majority of our uncertainty."},{"dimension":"media","status":"pass","explanation":"No media was selected and no verified media candidates were supplied, so there is no media identity or relevance issue.","excerpt":null}],"createdAt":"2026-10-29T15:56:45.326Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}