{"id":"5b2d50a4df77b8ff2f16a35c05569d2c3d0914f5a9993ce9a36adb63bd542cde","title":"Toby Ord argues hazard rates decline as AI-agent tasks continue","argument":"Toby Ord published this argument on February 4, 2026. He says new analysis by Gus Hamilton challenges his earlier constant hazard rate model: estimates using the Weibull distribution put its parameter k below 1 across all models, indicating that hazard rates decline as tasks continue. Ord says this changes estimates on either side of the 50% horizon: longer-task success rates are better, and shorter-task rates worse, than his earlier model suggested. He argues that comparisons between humans and AI agents depend on task length: with the same 50% horizon, humans’ lower k would mean better success on very long tasks but worse success on short ones. He also says k values for AI agents cluster around 0.6, while changes in λ explain improvements in METR timelines. Ord is puzzled that Opus 4.5’s estimated k resembles other AI agents despite its λ resembling humans, and notes that individual-model estimates may be noisy. His main caution is that the Weibull and log-logistic models differ substantially in their tail estimates, including the 99% horizon. He slightly favors Weibull, but says the evidence does not identify which model is right, leaving 99%-reliability horizons deeply uncertain—a threshold he says is necessary for much useful work.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 25, 2026; originally published February 4, 2026. Ord says estimates for individual models may be noisy and that there is not much data to assess Opus 4.5.","speaker":{"name":"Toby Ord","roleAtSource":"Publisher"},"source":{"title":"Hazard Rates for AI Agents Decline as a Task Goes On","publisher":"Toby Ord","format":"Article","url":"https://www.tobyord.com/writing/hazard-rates-for-ai-agents-decline","publishedAt":"2026-02-04T17:35:35+00:00","dateBasis":"Date stated on the publisher’s page."},"evidence":[{"text":"Toby Ord says Gus Hamilton’s analysis found decreasing hazard rates across all models using the Weibull distribution.","excerpt":"Gus found that the estimated values for k were below 1 for all the models, showing that all of them had decreasing hazard rates as the task went on.","url":"https://www.tobyord.com/writing/hazard-rates-for-ai-agents-decline","relationship":"Source-stated"},{"text":"Ord says the constant hazard rate model overestimates success on shorter tasks and underestimates it on tasks beyond the 50% horizon.","excerpt":"AI agents’ success rates on tasks beyond their 50%-horizon are better than my constant-hazard-rate model suggests, but those for tasks shorter than the 50% horizon are worse.","url":"https://www.tobyord.com/writing/hazard-rates-for-ai-agents-decline","relationship":"Source-stated"},{"text":"Ord says human and AI-agent success rates differ by task length even if their 50% horizons are the same.","excerpt":"This means even were they to have the same 50% horizon, humans would have a higher success rate on really long tasks (but a lower rate on really short ones).","url":"https://www.tobyord.com/writing/hazard-rates-for-ai-agents-decline","relationship":"Source-stated"},{"text":"Ord says it remains unknown which model best estimates the tails, leaving 99%-reliability horizons deeply uncertain.","excerpt":"the bigger lesson is that it is unknown which is right, and thus the 99%-reliability horizons (a threshold that is necessary for much useful work) are deeply uncertain.","url":"https://www.tobyord.com/writing/hazard-rates-for-ai-agents-decline","relationship":"Source-stated"},{"text":"Ord slightly favors the Weibull model, while noting that the log-logistic model fits slightly better.","excerpt":"He finds that they are similar, but with the log-logistic fitting slightly better.","url":"https://www.tobyord.com/writing/hazard-rates-for-ai-agents-decline","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.tobyord.com/writing/hazard-rates-for-ai-agents-decline","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What additional data would distinguish the Weibull and log-logistic models in their estimates of very short and very long task horizons?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Toby Ord argues hazard rates decline as AI-agent tasks continue","summary":"Toby Ord published this argument on February 4, 2026. He says new analysis by Gus Hamilton challenges his earlier constant hazard rate model: estimates using the Weibull distribution put its parameter k below 1 across all models, indicating that hazard rates decline as tasks continue. Ord says this changes estimates on either side of the 50% horizon: longer-task success rates are better, and shorter-task rates worse, than his earlier model suggested. He argues that comparisons between humans and AI agents depend on task length: with the same 50% horizon, humans’ lower k would mean better success on very long tasks but worse success on short ones. He also says k values for AI agents cluster around 0.6, while changes in λ explain improvements in METR timelines. Ord is puzzled that Opus 4.5’s estimated k resembles other AI agents despite its λ resembling humans, and notes that individual-model estimates may be noisy. His main caution is that the Weibull and log-logistic models differ substantially in their tail estimates, including the 99% horizon. He slightly favors Weibull, but says the evidence does not identify which model is right, leaving 99%-reliability horizons deeply uncertain—a threshold he says is necessary for much useful work.","question":"What additional data would distinguish the Weibull and log-logistic models in their estimates of very short and very long task horizons?","claims":[{"text":"Toby Ord says Gus Hamilton’s analysis found decreasing hazard rates across all models using the Weibull distribution.","excerpt":"Gus found that the estimated values for k were below 1 for all the models, showing that all of them had decreasing hazard rates as the task went on.","support":"source-stated"},{"text":"Ord says the constant hazard rate model overestimates success on shorter tasks and underestimates it on tasks beyond the 50% horizon.","excerpt":"AI agents’ success rates on tasks beyond their 50%-horizon are better than my constant-hazard-rate model suggests, but those for tasks shorter than the 50% horizon are worse.","support":"source-stated"},{"text":"Ord says human and AI-agent success rates differ by task length even if their 50% horizons are the same.","excerpt":"This means even were they to have the same 50% horizon, humans would have a higher success rate on really long tasks (but a lower rate on really short ones).","support":"source-stated"},{"text":"Ord says it remains unknown which model best estimates the tails, leaving 99%-reliability horizons deeply uncertain.","excerpt":"the bigger lesson is that it is unknown which is right, and thus the 99%-reliability horizons (a threshold that is necessary for much useful work) are deeply uncertain.","support":"source-stated"},{"text":"Ord slightly favors the Weibull model, while noting that the log-logistic model fits slightly better.","excerpt":"He finds that they are similar, but with the log-logistic fitting slightly better.","support":"source-stated"}],"limitations":["Text as retrieved on October 25, 2026; originally published February 4, 2026.","Ord says estimates for individual models may be noisy and that there is not much data to assess Opus 4.5."],"mediaId":null,"correctionReason":null,"id":"5b2d50a4df77b8ff2f16a35c05569d2c3d0914f5a9993ce9a36adb63bd542cde","publicationId":"30b74863-39b3-73c5-2c13-5459ce5d2c27","version":1,"sourceUrl":"https://www.tobyord.com/writing/hazard-rates-for-ai-agents-decline","sourceTitle":"Hazard Rates for AI Agents Decline as a Task Goes On","publisher":"Toby Ord","publishedAt":"2026-02-04T17:35:35+00:00","originalPublished":{"value":"2026-02-04T17:35:35+00:00","precision":"instant","basis":"backfill:article_published"},"firstObservedAt":"2026-10-25T19:35:45.168Z","createdAt":"2026-11-08T09:23:23.411Z","historical":true,"entityIds":[],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.tobyord.com/writing/hazard-rates-for-ai-agents-decline","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"2a4ead331c3c6860908df8218d4f66a2e1aed23bbf05e370b2bb749ab326e31b","rubricVersion":"unrecorded","provider":"vercel","model":"typesafe-ai/jev","reasonCodes":["bridge_admitted_by_model"],"confidence":0.62,"decidedAt":"2026-10-25T19:35:44.445Z"},"provenance":{"textHash":"fb443a12de0619014a3f749c0d1b79ec8684fac7eb1e5dfb9b4174d99eb10c75","observationId":"18e8092e39b907e4d7cf414a6531757322951ebe4ba4584861ba626f3a7786b7","retrievedAt":"2026-10-25T19:41:37.656Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"782d4421-c25e-96a0-0329-7594f5fd182a"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"0c3e6ed2-4674-3559-89fd-d2080dd8f08d"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-08T09:18:55.709Z","model":"anthropic/claude-opus-5.5"}},"canonicalUrl":"https://escapemeridian.com/readings/toby-ord-argues-hazard-rates-decline-as-ai-agent-tasks-continue-5b2d50a4","formats":{"html":"/readings/toby-ord-argues-hazard-rates-decline-as-ai-agent-tasks-continue-5b2d50a4","markdown":"/readings/toby-ord-argues-hazard-rates-decline-as-ai-agent-tasks-continue-5b2d50a4.md","json":"/readings/toby-ord-argues-hazard-rates-decline-as-ai-agent-tasks-continue-5b2d50a4.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-ai-agent","name":"AI agent","explanation":"An AI system that chooses and carries out actions toward a goal, often using software tools.","referenceUrl":"https://www.ibm.com/think/topics/ai-agents"}],"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":"5b2d50a4df77b8ff2f16a35c05569d2c3d0914f5a9993ce9a36adb63bd542cde","revision":"30b74863-39b3-73c5-2c13-5459ce5d2c27","revisionKind":"publication-id","title":"Toby Ord argues hazard rates decline as AI-agent tasks continue","url":"https://escapemeridian.com/readings/toby-ord-argues-hazard-rates-decline-as-ai-agent-tasks-continue-5b2d50a4"},"source":{"type":"source","id":"https://www.tobyord.com/writing/hazard-rates-for-ai-agents-decline","revision":"fb443a12de0619014a3f749c0d1b79ec8684fac7eb1e5dfb9b4174d99eb10c75","revisionKind":"sha256-source-text","title":"Hazard Rates for AI Agents Decline as a Task Goes On","url":"https://www.tobyord.com/writing/hazard-rates-for-ai-agents-decline"},"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":"5b2d50a4df77b8ff2f16a35c05569d2c3d0914f5a9993ce9a36adb63bd542cde","title":"Toby Ord argues hazard rates decline as AI-agent tasks continue","argument":"Toby Ord published this argument on February 4, 2026. He says new analysis by Gus Hamilton challenges his earlier constant hazard rate model: estimates using the Weibull distribution put its parameter k below 1 across all models, indicating that hazard rates decline as tasks continue. Ord says this changes estimates on either side of the 50% horizon: longer-task success rates are better, and shorter-task rates worse, than his earlier model suggested. He argues that comparisons between humans and AI agents depend on task length: with the same 50% horizon, humans’ lower k would mean better success on very long tasks but worse success on short ones. He also says k values for AI agents cluster around 0.6, while changes in λ explain improvements in METR timelines. Ord is puzzled that Opus 4.5’s estimated k resembles other AI agents despite its λ resembling humans, and notes that individual-model estimates may be noisy. His main caution is that the Weibull and log-logistic models differ substantially in their tail estimates, including the 99% horizon. He slightly favors Weibull, but says the evidence does not identify which model is right, leaving 99%-reliability horizons deeply uncertain—a threshold he says is necessary for much useful work.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 25, 2026; originally published February 4, 2026. Ord says estimates for individual models may be noisy and that there is not much data to assess Opus 4.5.","speaker":{"name":"Toby Ord","roleAtSource":"Publisher"},"source":{"title":"Hazard Rates for AI Agents Decline as a Task Goes On","publisher":"Toby Ord","format":"Article","url":"https://www.tobyord.com/writing/hazard-rates-for-ai-agents-decline","publishedAt":"2026-02-04T17:35:35+00:00","dateBasis":"Date stated on the publisher’s page."},"evidence":[{"text":"Toby Ord says Gus Hamilton’s analysis found decreasing hazard rates across all models using the Weibull distribution.","excerpt":"Gus found that the estimated values for k were below 1 for all the models, showing that all of them had decreasing hazard rates as the task went on.","url":"https://www.tobyord.com/writing/hazard-rates-for-ai-agents-decline","relationship":"Source-stated"},{"text":"Ord says the constant hazard rate model overestimates success on shorter tasks and underestimates it on tasks beyond the 50% horizon.","excerpt":"AI agents’ success rates on tasks beyond their 50%-horizon are better than my constant-hazard-rate model suggests, but those for tasks shorter than the 50% horizon are worse.","url":"https://www.tobyord.com/writing/hazard-rates-for-ai-agents-decline","relationship":"Source-stated"},{"text":"Ord says human and AI-agent success rates differ by task length even if their 50% horizons are the same.","excerpt":"This means even were they to have the same 50% horizon, humans would have a higher success rate on really long tasks (but a lower rate on really short ones).","url":"https://www.tobyord.com/writing/hazard-rates-for-ai-agents-decline","relationship":"Source-stated"},{"text":"Ord says it remains unknown which model best estimates the tails, leaving 99%-reliability horizons deeply uncertain.","excerpt":"the bigger lesson is that it is unknown which is right, and thus the 99%-reliability horizons (a threshold that is necessary for much useful work) are deeply uncertain.","url":"https://www.tobyord.com/writing/hazard-rates-for-ai-agents-decline","relationship":"Source-stated"},{"text":"Ord slightly favors the Weibull model, while noting that the log-logistic model fits slightly better.","excerpt":"He finds that they are similar, but with the log-logistic fitting slightly better.","url":"https://www.tobyord.com/writing/hazard-rates-for-ai-agents-decline","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.tobyord.com/writing/hazard-rates-for-ai-agents-decline","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What additional data would distinguish the Weibull and log-logistic models in their estimates of very short and very long task horizons?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Toby Ord argues hazard rates decline as AI-agent tasks continue","summary":"Toby Ord published this argument on February 4, 2026. He says new analysis by Gus Hamilton challenges his earlier constant hazard rate model: estimates using the Weibull distribution put its parameter k below 1 across all models, indicating that hazard rates decline as tasks continue. Ord says this changes estimates on either side of the 50% horizon: longer-task success rates are better, and shorter-task rates worse, than his earlier model suggested. He argues that comparisons between humans and AI agents depend on task length: with the same 50% horizon, humans’ lower k would mean better success on very long tasks but worse success on short ones. He also says k values for AI agents cluster around 0.6, while changes in λ explain improvements in METR timelines. Ord is puzzled that Opus 4.5’s estimated k resembles other AI agents despite its λ resembling humans, and notes that individual-model estimates may be noisy. His main caution is that the Weibull and log-logistic models differ substantially in their tail estimates, including the 99% horizon. He slightly favors Weibull, but says the evidence does not identify which model is right, leaving 99%-reliability horizons deeply uncertain—a threshold he says is necessary for much useful work.","question":"What additional data would distinguish the Weibull and log-logistic models in their estimates of very short and very long task horizons?","claims":[{"text":"Toby Ord says Gus Hamilton’s analysis found decreasing hazard rates across all models using the Weibull distribution.","excerpt":"Gus found that the estimated values for k were below 1 for all the models, showing that all of them had decreasing hazard rates as the task went on.","support":"source-stated"},{"text":"Ord says the constant hazard rate model overestimates success on shorter tasks and underestimates it on tasks beyond the 50% horizon.","excerpt":"AI agents’ success rates on tasks beyond their 50%-horizon are better than my constant-hazard-rate model suggests, but those for tasks shorter than the 50% horizon are worse.","support":"source-stated"},{"text":"Ord says human and AI-agent success rates differ by task length even if their 50% horizons are the same.","excerpt":"This means even were they to have the same 50% horizon, humans would have a higher success rate on really long tasks (but a lower rate on really short ones).","support":"source-stated"},{"text":"Ord says it remains unknown which model best estimates the tails, leaving 99%-reliability horizons deeply uncertain.","excerpt":"the bigger lesson is that it is unknown which is right, and thus the 99%-reliability horizons (a threshold that is necessary for much useful work) are deeply uncertain.","support":"source-stated"},{"text":"Ord slightly favors the Weibull model, while noting that the log-logistic model fits slightly better.","excerpt":"He finds that they are similar, but with the log-logistic fitting slightly better.","support":"source-stated"}],"limitations":["Text as retrieved on October 25, 2026; originally published February 4, 2026.","Ord says estimates for individual models may be noisy and that there is not much data to assess Opus 4.5."],"mediaId":null,"correctionReason":null,"id":"5b2d50a4df77b8ff2f16a35c05569d2c3d0914f5a9993ce9a36adb63bd542cde","publicationId":"30b74863-39b3-73c5-2c13-5459ce5d2c27","version":1,"sourceUrl":"https://www.tobyord.com/writing/hazard-rates-for-ai-agents-decline","sourceTitle":"Hazard Rates for AI Agents Decline as a Task Goes On","publisher":"Toby Ord","publishedAt":"2026-02-04T17:35:35+00:00","originalPublished":{"value":"2026-02-04T17:35:35+00:00","precision":"instant","basis":"backfill:article_published"},"firstObservedAt":"2026-10-25T19:35:45.168Z","createdAt":"2026-11-08T09:23:23.411Z","historical":true,"entityIds":[],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.tobyord.com/writing/hazard-rates-for-ai-agents-decline","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"2a4ead331c3c6860908df8218d4f66a2e1aed23bbf05e370b2bb749ab326e31b","rubricVersion":"unrecorded","provider":"vercel","model":"typesafe-ai/jev","reasonCodes":["bridge_admitted_by_model"],"confidence":0.62,"decidedAt":"2026-10-25T19:35:44.445Z"},"provenance":{"textHash":"fb443a12de0619014a3f749c0d1b79ec8684fac7eb1e5dfb9b4174d99eb10c75","observationId":"18e8092e39b907e4d7cf414a6531757322951ebe4ba4584861ba626f3a7786b7","retrievedAt":"2026-10-25T19:41:37.656Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"782d4421-c25e-96a0-0329-7594f5fd182a"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"0c3e6ed2-4674-3559-89fd-d2080dd8f08d"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-08T09:18:55.709Z","model":"anthropic/claude-opus-5.5"}}}],"audits":[{"publicationId":"30b74863-39b3-73c5-2c13-5459ce5d2c27","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading attributes the argument to Toby Ord and the underlying Weibull analysis to Gus Hamilton, matching the source. Ord's own views, such as slightly favoring Weibull and being puzzled about Opus 4.5, are kept distinct from Hamilton's findings.","excerpt":"But some excellent new analysis by Gus Hamilton provides this evidence."},{"dimension":"summary","status":"uncertain","explanation":"The summary is faithful, but central terms are unexplained for a general reader. Only 'AI agent' is defined. The reading never says in plain language what a hazard rate, the 50% horizon or 99% horizon, the Weibull parameter k or λ, or METR are, so the core claim that 'hazard rates decline' is hard to grasp. Smallest repair: plain-language glosses or term-library tooltips for hazard rate, 50%/99% horizon, k and METR. The question is substantive and fair.","excerpt":"It has a second parameter, k, which represents how the hazard rate changes with time (if at all)."},{"dimension":"date","status":"uncertain","explanation":"February 4, 2026 matches the UTC day of originalPublished and the page dateline, and no hindsight framing appears. However, editedAfterPublication is true: the page was modified 2026-03-18 per json-ld dateModified. Passages such as the Opus 4.5 'word of warning' cannot be confirmed as part of the original piece rather than a later edit.","excerpt":"February 4, 2026Toby Ord"},{"dimension":"support","status":"uncertain","explanation":"Each claim matches the source text: k below 1 for all models; better success beyond the 50% horizon and worse below it; the human-versus-AI comparison; Ord slightly favoring Weibull; and the deep uncertainty about 99% horizons. Because the page was edited after publication (modified 2026-03-18), it is not plain that every cited passage belongs to the original piece, notably the Opus 4.5 remarks. The reading also adds no unsupported benefits or costs.","excerpt":"I'd slightly favour the Weibull, on the grounds that it is more plausible ex ante."},{"dimension":"media","status":"pass","explanation":"No media was selected and no verified candidates were supplied, so nothing was invented or mismatched.","excerpt":null}],"createdAt":"2026-11-08T09:18:55.709Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}