{"id":"d67c5df6e761d86b32b85f1d5d5e1f7e34acc13544803aee52fc8331410c57dd","title":"Epoch AI argues AI performance is getting cheaper at an extraordinary pace","argument":"Epoch AI published this report on September 22, 2026. It argues that the cost of achieving a given level of AI benchmark performance has fallen about 47% per quarter, or 13× per year, across five benchmarks with data starting in 2023. The report compares this pace with historical price declines for other technologies, but cautions that “price of thought” bundles changing services and is not directly comparable to a single product such as electricity.\n\nEpoch AI estimates the cheapest available model able to reach each benchmark score. It finds that prices often fall fastest when a performance level first becomes state of the art (SOTA), then slow: the estimated drop averages 66% per quarter at SOTA and 32% two years later. The report suggests a possible explanation: a model maker may briefly charge a premium for newly achieved performance before competitors catch up. Epoch AI says fleeting supra-normal profits for a given AI service may be both a consequence and a cause of the industry’s race dynamic.\n\nThese estimates concern benchmark performance, not useful work or the experience of users who do not continually switch to the cheapest model. Possible benchmaxxing, incomplete and noisy data, and choices in averaging also limit the estimates.","topicIds":["investment-booms-and-busts","artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 23, 2026; originally published September 22, 2026. Epoch AI says benchmark performance is an imperfect proxy for useful work and market value, and that users who do not continually search for the cheapest model may not realize the estimated savings. The report flags possible benchmaxxing, incomplete and noisy data, a short measurement period, and choices in how results are averaged. It cautions that its estimates should not be read as exact. Epoch AI says “price of thought” bundles an evolving range of services, making comparisons with more homogeneous products such as electricity imperfect.","speaker":{"name":"Epoch AI","roleAtSource":"Publisher"},"source":{"title":"The plunging price of thought","publisher":"Epoch AI","format":"Article","url":"https://epoch.ai/publications/the-plunging-price-of-thought","publishedAt":"2026-09-22T00:00:00.000Z","dateBasis":"Date supplied by the publisher’s feed."},"evidence":[{"text":"Epoch AI estimates that the cost of a given level of AI performance fell about 47% per quarter, or 13× per year, across its five primary benchmarks.","excerpt":"Across that time, we find that the price for a given level of performance has fallen about 47% per quarter, or 13× per year.","url":"https://epoch.ai/publications/the-plunging-price-of-thought","relationship":"Source-stated"},{"text":"Epoch AI estimates faster price declines near state-of-the-art performance, with the pace slowing over the following two years.","excerpt":"On average for the primary benchmarks, the decline rate drops from 66% per quarter on average at SOTA to 32% two years later.","url":"https://epoch.ai/publications/the-plunging-price-of-thought","relationship":"Source-stated"},{"text":"Epoch AI says benchmark results may not reflect useful work, and that users who do not continually seek the cheapest model may not experience the same savings.","excerpt":"Essentially no user stays permanently on the cost frontier, checking all available models to find which can most cheaply execute each task up to a well-defined level of performance.","url":"https://epoch.ai/publications/the-plunging-price-of-thought","relationship":"Source-stated"},{"text":"Epoch AI suggests that temporary premiums for newly achieved performance, followed by competitors catching up, could help explain why price declines are fastest near SOTA.","excerpt":"One scenario that would explain this pattern: when a new, SOTA-defining model appears, its maker can briefly charge a premium for this performance.\nBut competitors open and closed quickly catch up, and the price of such performance plunges.","url":"https://epoch.ai/publications/the-plunging-price-of-thought","relationship":"Source-stated"},{"text":"Epoch AI says the fleeting nature of supra-normal profits for a given AI service may be both a consequence and a cause of the industry's race dynamic.","excerpt":"This fact is surely both a consequence and a cause of the race dynamic in the AI industry today.","url":"https://epoch.ai/publications/the-plunging-price-of-thought","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://epoch.ai/publications/the-plunging-price-of-thought","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How closely do these benchmark price declines track the cost of AI services users actually find useful?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Epoch AI argues AI performance is getting cheaper at an extraordinary pace","summary":"Epoch AI published this report on September 22, 2026. It argues that the cost of achieving a given level of AI benchmark performance has fallen about 47% per quarter, or 13× per year, across five benchmarks with data starting in 2023. The report compares this pace with historical price declines for other technologies, but cautions that “price of thought” bundles changing services and is not directly comparable to a single product such as electricity.\n\nEpoch AI estimates the cheapest available model able to reach each benchmark score. It finds that prices often fall fastest when a performance level first becomes state of the art (SOTA), then slow: the estimated drop averages 66% per quarter at SOTA and 32% two years later. The report suggests a possible explanation: a model maker may briefly charge a premium for newly achieved performance before competitors catch up. Epoch AI says fleeting supra-normal profits for a given AI service may be both a consequence and a cause of the industry’s race dynamic.\n\nThese estimates concern benchmark performance, not useful work or the experience of users who do not continually switch to the cheapest model. Possible benchmaxxing, incomplete and noisy data, and choices in averaging also limit the estimates.","question":"How closely do these benchmark price declines track the cost of AI services users actually find useful?","claims":[{"text":"Epoch AI estimates that the cost of a given level of AI performance fell about 47% per quarter, or 13× per year, across its five primary benchmarks.","excerpt":"Across that time, we find that the price for a given level of performance has fallen about 47% per quarter, or 13× per year.","support":"source-stated"},{"text":"Epoch AI estimates faster price declines near state-of-the-art performance, with the pace slowing over the following two years.","excerpt":"On average for the primary benchmarks, the decline rate drops from 66% per quarter on average at SOTA to 32% two years later.","support":"source-stated"},{"text":"Epoch AI says benchmark results may not reflect useful work, and that users who do not continually seek the cheapest model may not experience the same savings.","excerpt":"Essentially no user stays permanently on the cost frontier, checking all available models to find which can most cheaply execute each task up to a well-defined level of performance.","support":"source-stated"},{"text":"Epoch AI suggests that temporary premiums for newly achieved performance, followed by competitors catching up, could help explain why price declines are fastest near SOTA.","excerpt":"One scenario that would explain this pattern: when a new, SOTA-defining model appears, its maker can briefly charge a premium for this performance.\nBut competitors open and closed quickly catch up, and the price of such performance plunges.","support":"source-stated"},{"text":"Epoch AI says the fleeting nature of supra-normal profits for a given AI service may be both a consequence and a cause of the industry's race dynamic.","excerpt":"This fact is surely both a consequence and a cause of the race dynamic in the AI industry today.","support":"source-stated"}],"limitations":["Text as retrieved on October 23, 2026; originally published September 22, 2026.","Epoch AI says benchmark performance is an imperfect proxy for useful work and market value, and that users who do not continually search for the cheapest model may not realize the estimated savings.","The report flags possible benchmaxxing, incomplete and noisy data, a short measurement period, and choices in how results are averaged. It cautions that its estimates should not be read as exact.","Epoch AI says “price of thought” bundles an evolving range of services, making comparisons with more homogeneous products such as electricity imperfect."],"mediaId":null,"correctionReason":null,"id":"d67c5df6e761d86b32b85f1d5d5e1f7e34acc13544803aee52fc8331410c57dd","publicationId":"70efac68-326b-2e7a-f357-82ea84898cb7","version":1,"sourceUrl":"https://epoch.ai/publications/the-plunging-price-of-thought","sourceTitle":"The plunging price of thought","publisher":"Epoch AI","publishedAt":"2026-09-22T00:00:00.000Z","originalPublished":{"value":"2026-09-22T00:00:00.000Z","precision":"instant","basis":"backfill:feed_published"},"firstObservedAt":"2026-10-18T02:45:42.839Z","createdAt":"2026-11-05T17:13:24.248Z","historical":true,"entityIds":["epoch-ai"],"topicIds":["investment-booms-and-busts","artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://epoch.ai/publications/the-plunging-price-of-thought","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"fe496b0a918df3b3f772294b630487eac2459665ab390b291b6a484d820d5d6a","rubricVersion":"unrecorded","provider":"unrecorded","model":"anthropic/claude-opus-5.5","reasonCodes":["bridge_admitted_by_model","second_opinion_categorical"],"confidence":1,"decidedAt":"2026-10-23T20:39:48.428Z"},"provenance":{"textHash":"86b2f8b7bc5485b3077519ec6e5db5c5b6d7588dc0d9a6daf249000e66168a84","observationId":"1cf15d80bf6846e54821edbea5a5bb35135d01ee386e7f7cf3603fb828a346e6","retrievedAt":"2026-10-23T20:41:41.586Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"dac307f2-f857-54de-9045-7c620a91a874"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"86a65e73-6568-be3e-f8ab-d96e1df8c55a"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-05T17:08:45.432Z","model":"anthropic/claude-opus-5.5"}},"canonicalUrl":"https://escapemeridian.com/readings/epoch-ai-argues-ai-performance-is-getting-cheaper-at-an-d67c5df6","formats":{"html":"/readings/epoch-ai-argues-ai-performance-is-getting-cheaper-at-an-d67c5df6","markdown":"/readings/epoch-ai-argues-ai-performance-is-getting-cheaper-at-an-d67c5df6.md","json":"/readings/epoch-ai-argues-ai-performance-is-getting-cheaper-at-an-d67c5df6.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-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":"epoch-ai","name":"Epoch AI","kind":"organization","context":"Epoch AI studies trends in AI capability, training resources and the infrastructure behind them.","referenceUrls":["https://epoch.ai/","https://epoch.ai/data"],"path":"/entities/epoch-ai","json":"/entities/epoch-ai.json","markdown":"/entities/epoch-ai.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"d67c5df6e761d86b32b85f1d5d5e1f7e34acc13544803aee52fc8331410c57dd","revision":"70efac68-326b-2e7a-f357-82ea84898cb7","revisionKind":"publication-id","title":"Epoch AI argues AI performance is getting cheaper at an extraordinary pace","url":"https://escapemeridian.com/readings/epoch-ai-argues-ai-performance-is-getting-cheaper-at-an-d67c5df6"},"source":{"type":"source","id":"https://epoch.ai/publications/the-plunging-price-of-thought","revision":"86b2f8b7bc5485b3077519ec6e5db5c5b6d7588dc0d9a6daf249000e66168a84","revisionKind":"sha256-source-text","title":"The plunging price of thought","url":"https://epoch.ai/publications/the-plunging-price-of-thought"},"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":"d67c5df6e761d86b32b85f1d5d5e1f7e34acc13544803aee52fc8331410c57dd","title":"Epoch AI argues AI performance is getting cheaper at an extraordinary pace","argument":"Epoch AI published this report on September 22, 2026. It argues that the cost of achieving a given level of AI benchmark performance has fallen about 47% per quarter, or 13× per year, across five benchmarks with data starting in 2023. The report compares this pace with historical price declines for other technologies, but cautions that “price of thought” bundles changing services and is not directly comparable to a single product such as electricity.\n\nEpoch AI estimates the cheapest available model able to reach each benchmark score. It finds that prices often fall fastest when a performance level first becomes state of the art (SOTA), then slow: the estimated drop averages 66% per quarter at SOTA and 32% two years later. The report suggests a possible explanation: a model maker may briefly charge a premium for newly achieved performance before competitors catch up. Epoch AI says fleeting supra-normal profits for a given AI service may be both a consequence and a cause of the industry’s race dynamic.\n\nThese estimates concern benchmark performance, not useful work or the experience of users who do not continually switch to the cheapest model. Possible benchmaxxing, incomplete and noisy data, and choices in averaging also limit the estimates.","topicIds":["investment-booms-and-busts","artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 23, 2026; originally published September 22, 2026. Epoch AI says benchmark performance is an imperfect proxy for useful work and market value, and that users who do not continually search for the cheapest model may not realize the estimated savings. The report flags possible benchmaxxing, incomplete and noisy data, a short measurement period, and choices in how results are averaged. It cautions that its estimates should not be read as exact. Epoch AI says “price of thought” bundles an evolving range of services, making comparisons with more homogeneous products such as electricity imperfect.","speaker":{"name":"Epoch AI","roleAtSource":"Publisher"},"source":{"title":"The plunging price of thought","publisher":"Epoch AI","format":"Article","url":"https://epoch.ai/publications/the-plunging-price-of-thought","publishedAt":"2026-09-22T00:00:00.000Z","dateBasis":"Date supplied by the publisher’s feed."},"evidence":[{"text":"Epoch AI estimates that the cost of a given level of AI performance fell about 47% per quarter, or 13× per year, across its five primary benchmarks.","excerpt":"Across that time, we find that the price for a given level of performance has fallen about 47% per quarter, or 13× per year.","url":"https://epoch.ai/publications/the-plunging-price-of-thought","relationship":"Source-stated"},{"text":"Epoch AI estimates faster price declines near state-of-the-art performance, with the pace slowing over the following two years.","excerpt":"On average for the primary benchmarks, the decline rate drops from 66% per quarter on average at SOTA to 32% two years later.","url":"https://epoch.ai/publications/the-plunging-price-of-thought","relationship":"Source-stated"},{"text":"Epoch AI says benchmark results may not reflect useful work, and that users who do not continually seek the cheapest model may not experience the same savings.","excerpt":"Essentially no user stays permanently on the cost frontier, checking all available models to find which can most cheaply execute each task up to a well-defined level of performance.","url":"https://epoch.ai/publications/the-plunging-price-of-thought","relationship":"Source-stated"},{"text":"Epoch AI suggests that temporary premiums for newly achieved performance, followed by competitors catching up, could help explain why price declines are fastest near SOTA.","excerpt":"One scenario that would explain this pattern: when a new, SOTA-defining model appears, its maker can briefly charge a premium for this performance.\nBut competitors open and closed quickly catch up, and the price of such performance plunges.","url":"https://epoch.ai/publications/the-plunging-price-of-thought","relationship":"Source-stated"},{"text":"Epoch AI says the fleeting nature of supra-normal profits for a given AI service may be both a consequence and a cause of the industry's race dynamic.","excerpt":"This fact is surely both a consequence and a cause of the race dynamic in the AI industry today.","url":"https://epoch.ai/publications/the-plunging-price-of-thought","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://epoch.ai/publications/the-plunging-price-of-thought","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How closely do these benchmark price declines track the cost of AI services users actually find useful?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Epoch AI argues AI performance is getting cheaper at an extraordinary pace","summary":"Epoch AI published this report on September 22, 2026. It argues that the cost of achieving a given level of AI benchmark performance has fallen about 47% per quarter, or 13× per year, across five benchmarks with data starting in 2023. The report compares this pace with historical price declines for other technologies, but cautions that “price of thought” bundles changing services and is not directly comparable to a single product such as electricity.\n\nEpoch AI estimates the cheapest available model able to reach each benchmark score. It finds that prices often fall fastest when a performance level first becomes state of the art (SOTA), then slow: the estimated drop averages 66% per quarter at SOTA and 32% two years later. The report suggests a possible explanation: a model maker may briefly charge a premium for newly achieved performance before competitors catch up. Epoch AI says fleeting supra-normal profits for a given AI service may be both a consequence and a cause of the industry’s race dynamic.\n\nThese estimates concern benchmark performance, not useful work or the experience of users who do not continually switch to the cheapest model. Possible benchmaxxing, incomplete and noisy data, and choices in averaging also limit the estimates.","question":"How closely do these benchmark price declines track the cost of AI services users actually find useful?","claims":[{"text":"Epoch AI estimates that the cost of a given level of AI performance fell about 47% per quarter, or 13× per year, across its five primary benchmarks.","excerpt":"Across that time, we find that the price for a given level of performance has fallen about 47% per quarter, or 13× per year.","support":"source-stated"},{"text":"Epoch AI estimates faster price declines near state-of-the-art performance, with the pace slowing over the following two years.","excerpt":"On average for the primary benchmarks, the decline rate drops from 66% per quarter on average at SOTA to 32% two years later.","support":"source-stated"},{"text":"Epoch AI says benchmark results may not reflect useful work, and that users who do not continually seek the cheapest model may not experience the same savings.","excerpt":"Essentially no user stays permanently on the cost frontier, checking all available models to find which can most cheaply execute each task up to a well-defined level of performance.","support":"source-stated"},{"text":"Epoch AI suggests that temporary premiums for newly achieved performance, followed by competitors catching up, could help explain why price declines are fastest near SOTA.","excerpt":"One scenario that would explain this pattern: when a new, SOTA-defining model appears, its maker can briefly charge a premium for this performance.\nBut competitors open and closed quickly catch up, and the price of such performance plunges.","support":"source-stated"},{"text":"Epoch AI says the fleeting nature of supra-normal profits for a given AI service may be both a consequence and a cause of the industry's race dynamic.","excerpt":"This fact is surely both a consequence and a cause of the race dynamic in the AI industry today.","support":"source-stated"}],"limitations":["Text as retrieved on October 23, 2026; originally published September 22, 2026.","Epoch AI says benchmark performance is an imperfect proxy for useful work and market value, and that users who do not continually search for the cheapest model may not realize the estimated savings.","The report flags possible benchmaxxing, incomplete and noisy data, a short measurement period, and choices in how results are averaged. It cautions that its estimates should not be read as exact.","Epoch AI says “price of thought” bundles an evolving range of services, making comparisons with more homogeneous products such as electricity imperfect."],"mediaId":null,"correctionReason":null,"id":"d67c5df6e761d86b32b85f1d5d5e1f7e34acc13544803aee52fc8331410c57dd","publicationId":"70efac68-326b-2e7a-f357-82ea84898cb7","version":1,"sourceUrl":"https://epoch.ai/publications/the-plunging-price-of-thought","sourceTitle":"The plunging price of thought","publisher":"Epoch AI","publishedAt":"2026-09-22T00:00:00.000Z","originalPublished":{"value":"2026-09-22T00:00:00.000Z","precision":"instant","basis":"backfill:feed_published"},"firstObservedAt":"2026-10-18T02:45:42.839Z","createdAt":"2026-11-05T17:13:24.248Z","historical":true,"entityIds":["epoch-ai"],"topicIds":["investment-booms-and-busts","artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://epoch.ai/publications/the-plunging-price-of-thought","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"fe496b0a918df3b3f772294b630487eac2459665ab390b291b6a484d820d5d6a","rubricVersion":"unrecorded","provider":"unrecorded","model":"anthropic/claude-opus-5.5","reasonCodes":["bridge_admitted_by_model","second_opinion_categorical"],"confidence":1,"decidedAt":"2026-10-23T20:39:48.428Z"},"provenance":{"textHash":"86b2f8b7bc5485b3077519ec6e5db5c5b6d7588dc0d9a6daf249000e66168a84","observationId":"1cf15d80bf6846e54821edbea5a5bb35135d01ee386e7f7cf3603fb828a346e6","retrievedAt":"2026-10-23T20:41:41.586Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"dac307f2-f857-54de-9045-7c620a91a874"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"86a65e73-6568-be3e-f8ab-d96e1df8c55a"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-05T17:08:45.432Z","model":"anthropic/claude-opus-5.5"}}}],"audits":[{"publicationId":"70efac68-326b-2e7a-f357-82ea84898cb7","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading is a single Epoch AI report, and it attributes the estimates, the possible explanation and the caveats to Epoch AI rather than presenting them as established fact.","excerpt":"While the fundamental finding of rapid price decline is indisputable, our particular estimates are subject to major caveats."},{"dimension":"summary","status":"uncertain","explanation":"The reading flattens the report's central thesis. It says only that Epoch compares the pace with other technologies, leaving out the claim that AI got cheaper faster than any other transformative technology in history (for example, 54 times faster than electricity). Smallest repair: state that thesis as Epoch's claim, with its caveat. The term benchmaxxing is also unexplained and should get a term-library tooltip.","excerpt":"AI has gotten cheaper more quickly than any other transformative technology in history."},{"dimension":"date","status":"pass","explanation":"The reading's September 22, 2026 date matches the UTC day of originalPublished. Its internal references, including a March 2026 paper and a model released under 18 months after January 31, 2025, fit that date. There is no hindsight framing, and the evidence is a publisher page, not an archive copy.","excerpt":null},{"dimension":"support","status":"uncertain","explanation":"Claim 5 and the summary weaken the source's conviction. The source hedges only on whether profits are fleeting ('may be fleeting'), then says this is 'surely' both a consequence and a cause of the race dynamic. The reading instead says the fleeting profits 'may be' both. Smallest repair: keep 'may be fleeting' and restore 'surely' for the consequence-and-cause link. The other claims match their excerpts.","excerpt":"This fact is surely both a consequence and a cause of the race dynamic in the AI industry today."},{"dimension":"media","status":"pass","explanation":"No media is selected, and no verified candidates were supplied, so nothing is misidentified.","excerpt":null}],"createdAt":"2026-11-05T17:08:45.432Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}