SOURCESStart with the source
The evidence behind our model comparisons, company analysis and country profiles.
Research definitions and attribution
Charts are selected, reformatted visualizations of the cited data. Units, observation periods, population definitions and uncertainty remain source-specific. Where marked, data is reused under CC BY 4.0; source-specific licence terms remain applicable.
Model registry and coverage audit
Snapshot 2026-10-26. We retain 3,636 Epoch research entries, 467 OpenRouter API offerings and 7 reviewed first-party Mistral entries as distinct source observations. These counts overlap and must not be added as a global model total.
7 identities have reviewed cross-source links. 177 similar-name groups remain review candidates. Missing release dates stay unknown; API creation times are not release dates. API availability and prices are separate from development-country counts.
Country counts use dated Epoch entries from 2010 onwards. Multinational models can appear in several countries, while the EU27 aggregate counts each source entry once. Current-year observations are partial. Official documentation may include retired models and hosted third-party models; presence on a provider website does not establish authorship. Family grouping does not merge versions, sizes or API routes into new releases.
Epoch AI · all model entries · captured 2026-10-26 · SHA-256 b8d55512966b52e09474abe6c8326aa73d5bb41e038b5a383b94f93c41951ade
OpenRouter · full API catalogue · captured 2026-10-25 · SHA-256 39b5c461aff3eee4d5d0cc8abe857e97cbadbf2274bde7ba4ac846da4a95c99a
Mistral · reviewed first-party selection · captured 2026-10-25 · SHA-256 436af2fc557411e93da8115e6a2a884a38c06e5d69b362d1dbd118a40459f30d
Country model projects · primary-source checks
Selected first-party checks of model projects and families. Not exhaustive; not additional model counts. Development attribution is distinct from training location or hosting. Project families and variants must not be summed with Epoch entries. Reviewed 2026-10-25.
Mistral model families · FR · Mistral AI
Pharia-1-LLM · DE · Aleph Alpha Research
Salamandra · ES · Barcelona Supercomputing Center / Aina
PLLuM · PL · Polish research consortium / Ministry of Digital Affairs
Apertus · CH · ETH Zurich / EPFL / CSCS
Sarvam 30B / 105B · IN · Sarvam AI
Additional primary-source model checks retain versioned cards and developer-country evidence in country profiles, including Slovakia, Czechia and Poland. These checks do not expand Epoch totals. Unknown lineage or release dates remain unestablished.
Epoch data: CC BY 4.0. Other source terms remain applicable. Official pages support selected factual checks; no blanket reuse licence is granted. This register is not exhaustive.
US business AI adoption by employment size
US employer businesses, by employment size. AI use in any business function in the last two weeks. Source-published four-period moving average of biweekly estimates; not raw biweekly observations.
Question broadened from producing goods or services to any business function in November 2025. Earlier wording is excluded. Dated July 2026 publication, not the latest Census release. No standard errors supplied in this figure. Size groups are not additive. Use does not measure productivity or implementation depth.
Published 2026-07-17 · retrieved and reviewed 2026-10-25. Figure_8, rows 5–17. Source workbook SHA-256: 97c331423b71ae5ab97962c165cbb192d67402f0b2024fb9194ecf52d2bb43d4
Federal Reserve research note and methodology · Official data workbook · Original Census BTOS seriesUS government statistical data. Federal Reserve authors’ analysis does not imply endorsement by the Board. Statvero preserves source-published smoothing and does not infer statistical significance without standard errors.
Claude usage
Published aggregate classifier estimates; Claude chat/Cowork and first-party API are different populations. API excludes Claude Code. Periods end exclusively; missing cells are unavailable.
CC BY · Anthropic Economic Index · captured 2026-10-24
CC BY · Anthropic Economic Index · captured 2026-10-24
API excludes Claude Code. Chat/Cowork and API have different populations. Classifier estimates and potentially overlapping categories are not workforce adoption or measured economic output. Source date_end is exclusive; missing and suppressed cells are not zero. Anthropic licenses data under CC BY and code under MIT.
LLM and robot exposure
733 exact occupation-code matches. LLM index 0–1 (March 2026) and robot index 0–3 (September 2026) have different definitions. Neither is a job-loss probability.
CC BY · Anthropic Economic Index · captured 2026-10-24
CC BY · Anthropic Economic Index · captured 2026-10-24
AI infrastructure
95 source-covered sites, non-exhaustive and geographically uneven. Current estimates as retrieved 24 October 2026. IT power capacity is not measured electricity consumption. Source-estimated shipments, grouped only by identical periods and manufacturer. Incomplete periods and future plans are retained; marginal uncertainty intervals are not summed.
CC BY 4.0 · Epoch AI · captured 2026-10-24
CC BY 4.0 · Epoch AI · captured 2026-10-24
Current IT power capacity is not measured electricity consumption. Coverage is geographically uneven and non-exhaustive. Future estimates, incomplete periods and observed snapshots are distinct. Individual uncertainty quantiles are not summed.
Stack Overflow Developer Survey
29 May–23 June 2025. Each survey question has its own response population; published figures are rounded.
AI assistance is not the same as vibe coding: Separate questions and response populations. These bars are not parts of one whole. Vibe coding means generating software from LLM prompts in professional work.
Use or plan to use AI: 33,662 responses; Vibe coding not part of professional work: 26,564 responses
More distrust than trust: Same accuracy question; neutral and other responses omitted. Do not normalize these two answers to 100%.
Distrust AI output accuracy: 33,244 responses; Trust AI output accuracy: 33,244 responses
Almost right still needs work: Multiple responses allowed. These percentages overlap; they must not be added. Perceptions are not measured debugging time.
Solutions almost right, but not quite: 31,476 responses; Debugging AI code takes more time: 31,476 responses
Guardrails and learning
Randomized field experiment · Nearly 1,000 high-school students · 2025-06-25
Practice grades and later unaided exam grades are separate outcomes; guarded-tutor exam effect is not encoded as zero.
Primary sourceAI support for human tutors
Randomized controlled trial · More than 700 tutors and 1,000 students · 2025-11
Tutor assistance, not replacement; low-rated subgroup is nested inside the full sample. Earlier paper versions used different samples.
Primary sourceCourse-integrated AI tutoring
Randomized controlled trial · 2,379 undergraduates; 30 instructors in the overall trial · 2026-10
Reported estimates are for comparisons among sections of the same course, not all students pooled. Working paper.
Primary sourceAnatomy learning across teaching approaches
Three-arm class-level randomized trial · 362 enrolled; 301 completed; six classes · 2026-05-22
16-week course; complete-case analysis; AI plus knowledge graphs is a teaching package, not an isolated AI effect. Reported ± values are SD, not confidence intervals.
Primary sourceAI-supported mammography screening
Randomized screening accuracy trial · 105,915 analysed women; four screening sites · 2025-03
Screening outcomes and reading workload, not survival or mortality; different arm sizes. Data collected April 2021–December 2022.
Primary sourceAI inside businesses
Multi-country executive survey · Nearly 6,000 senior executives; surveys November 2025–January 2026 · 2026-03
Self-reported use and effects, not a causal experiment. Country survey frames differ. Expectations are not observed results. Applications overlap. © 2026 Bank of England; selected factual summary.
Primary sourceAI coding assistance in ordinary work
Three company-run randomized field experiments · 4,867 developers across Microsoft, Accenture and an anonymous Fortune 100 company · 2025-06
Pooled result across three noisy company experiments; completed tasks are not code quality. Some authors worked for the tool vendor. Less experience is not a quantified subgroup effect here.
Primary sourceConsultants on both sides of the AI frontier
Preregistered randomized field experiment · 758 BCG consultants; three experimental conditions · 2026-03-11
Historical GPT-4, bounded consulting tasks; not economy-wide productivity. Tasks inside the model capability frontier and one task outside it use distinct outcomes. Organization Science, © 2026 The Authors, CC BY 4.0.
Primary sourceMASAI interval cancers and screening performance
Randomized screening trial · follow-up of the same MASAI cohort · 105,915 analysed women; same MASAI cohort, not a second independent trial · 2026-01-31
Interval cancers diagnosed between rounds or within two years of the last negative screen. Non-inferiority is not superiority: interval-cancer ratio 0.88, 95% CI 0.65–1.18. No mortality estimate. Enrollment April 2021–December 2022.
Primary sourceEducation gaps with AI assistance
Randomized online experiment · 1,174 adults aged 25–45 · 2026-10
Education-group performance gaps on one business problem-solving task, not national wages or long-run productivity. Follow-up gains are not assigned numerical values from the abstract.
Primary sourceAI and time spent on email
Randomized workplace field experiment · 7,137 knowledge workers across 66 firms · 2025-11
Two-hour email reduction concerns tool users (80% of treated workers), in the second half of a six-month experiment. Not the all-assigned-worker effect or an estimate of total productivity. Some authors worked for Microsoft.
Primary sourceWriting faster is not the same as replacing a job
453 college-educated professionals · Writing experiment; collection dates not specified in the reviewed abstract · published 2023-07-13
Short writing assignments, not complete jobs or economy-wide output. Experimental tasks required less factual and organizational context than many real assignments.
Limited independently written factual summary; Science default licence. No source figure, full table or participant data reproduced.
Primary study or reportCustomer support saw a measurable productivity gain
5,172 customer-support agents · Staggered workplace rollout studied in the 2025 journal edition; not a 2025 annual statistic · published 2025-02-04
One company and one workflow; not a randomized global productivity estimate. Use the 2025 journal result, not the 14% estimate from the earlier working-paper edition.
Limited factual summary with attribution to Brynjolfsson, Li and Raymond; no source figure or microdata reproduced.
Primary study or reportPerceived speed and measured speed diverged
16 experienced developers; 246 tasks · February–June 2025 · published 2025-07-10
Historical tools and a narrow sample; the reported 95% confidence interval is +2% to +39% task time. METR's February 2026 update warns that later experiments suffer from selection and time-measurement problems; this is not an estimate of current tool performance.
Limited factual summary credited to Becker, Rush, Barnes and Rein / METR; no figure or task data copied.
Primary study or reportA purpose-built tutor improved immediate test results
194 eligible students; repeated lesson observations, not 316 unique students · Fall 2023 semester; weeks 9–10; two Harvard undergraduate physics lessons · published 2025-06-03
Immediate outcomes in two lessons, not long-term retention or all education. Post-test observations: tutor n=142, class n=174. Classroom learning time is an assumed 60 minutes; tutor time is a measured median.
Limited factual summary of Kestin et al. Article: CC BY-NC-ND 4.0; no source artwork, text, full table or raw dataset is redistributed.
Primary study or reportA strong model did not establish a physician benefit
50 physicians; 244 completed vignette cases · 2023-11-29 → 2023-12-29 · published 2024-10-28
Rubric scores on written diagnostic cases, not patient outcomes or real-world diagnostic accuracy. The exploratory standalone score is not a randomized physician group. Adjusted difference: 2 pp, 95% CI −4 to 8; P=.60.
Goh et al., JAMA Network Open (2024), CC BY; adapted numeric summary, no patient data.
Primary study or reportData-centre demand has a growing physical footprint
Global data centres; AI-focused facilities are a subset · IEA April 2026 edition; 2025 estimate and 2030 projection · published 2026-04-16
Total data-centre consumption is not AI-only consumption. The 2030 value is a projection, not a measurement; no annual path is interpolated. Growth rates compare overlapping populations and are not additive.
IEA (2026), Key Questions on Energy and AI, IEA, Paris; CC BY 4.0. Adapted factual chart.
Primary study or report