AI Development Index: How AI Grew, 1950–2026

Home / Indexes / State of AI / AI Development Index
Report AI · Index · 1950–2026 · Updated quarterly

AI Development Index: How AI Grew, 1950–2026

A year-by-year record of how artificial intelligence developed, in six measures: training compute, capability, cost, capital, adoption and risk. Every value is dated, sourced and confidence-rated, and the full dataset is free to download.

Thank you for reading this post, don't forget to subscribe!

By the Report AI editorial team · Published · Cadence: quarterly (STANDARD) · Next update February 2027
~9.4×10²⁶
FLOP, the largest estimated training run to date (GPT-6 Astra, Epoch AI, Sep 2026) HIGH
5×
a year: Epoch AI’s growth rate for frontier language-model training compute since 2020 HIGH
$0.20
per million tokens: cheapest GPT-4-level price, Oct 2026 ($37.50 in Mar 2023) DERIVED
89%
of McKinsey survey respondents say their organisation uses AI in at least one function (2026) MEDIUM
At a glance
  • The largest estimated AI training run grew from about 40 operations for Claude Shannon’s maze-solving mouse in 1950 to about 9.4×10²⁶ for OpenAI’s GPT-6 Astra in September 2026, according to Epoch AI.
  • Epoch AI estimates frontier language-model training compute has grown about 5× a year since 2020, and the cost of those runs about 3.5× a year.
  • Capability tests are saturating within one to two years of launch: SWE-bench Verified went from 49% to 97.0% between late 2024 and September 2026.
  • The cheapest GPT-4-level model cost $37.50 per million tokens in March 2023 and about $0.20 in October 2026.
  • Global corporate AI investment, in constant 2021 dollars, rose from about $17 billion in 2013 to $490 billion in 2025 (Our World in Data, from the Stanford HAI AI Index).
  • Reported AI incidents in the AI Incident Database rose from 233 in 2024 to 362 in 2025, as cited by the AI Index 2026.
Cite this report
Report AI, AI Development Index: How AI Grew, 1950–2026, November 2026 update, published 16 November 2026. https://report-ai.org/indexes/state-of-ai/ai-development-index/
Download the data (CSV) · every table and exhibit on this page, with sources and confidence ratings. Free to quote with attribution and a link.
In this index
  1. Decade by decade — Every decade since 1950 raised the record for the largest training run except one: the 1970s, the decade of the first AI winter
  2. 1. Compute — Training compute: from 10⁶ FLOP for the Perceptron to about 10²⁷ for today’s frontier
  3. 2. Capability — Capability: benchmarks rise from near zero to saturation in one to five years
  4. 3. Cost — Cost: buying a fixed capability gets cheaper every year, while building the frontier gets dearer
  5. 4. Capital — Capital: corporate AI investment rose about 29-fold in real terms from 2013 to 2025
  6. 5. Adoption and research output — Adoption: from a specialist tool to a billion weekly users
  7. 6. Risk and resources — Risk and resources: incidents and electricity use are rising, but are the least complete series
  8. Register — The milestone register: 70 dated events from 1943 to 2026
  9. FAQ
  10. Definitions
  11. About the data
Analysis

This index tracks AI’s development through six measures that can be dated and sourced across decades, rather than through a single composite score. Composite indices hide their weights; separate measures let readers see that capability, cost and adoption have moved at very different speeds. The companion report, How AI Developed, 1950–2026, tells the story; this page holds the numbers.

The clearest long-run signal is compute. Epoch AI’s database of notable models shows the largest training run of each year rising by about 25 orders of magnitude since 1950, with a sharp change of pace around 2010: before then compute doubled roughly every 20 months, in line with general computing, and afterwards about every six months. Since 2020 the largest language-model runs have grown about 5× a year. The record in November 2026 is Epoch’s estimate of about 9.4×10²⁶ FLOP for GPT-6 Astra, roughly 45 times GPT-4.

Cost and capability have moved in opposite directions for buyers and builders. For buyers, a fixed level of capability gets dramatically cheaper: GPT-4-level output fell about 187-fold in price between March 2023 and October 2026. For builders, the frontier gets dearer: Epoch’s amortised estimate for the final training run of Grok 4 is about $390 million, against about $16 for AlexNet in 2012. Capital has followed the builders, with corporate AI investment rising about 29-fold in real terms from 2013 to 2025.

Adoption and risk measures are younger and less consistent. McKinsey’s survey shows organisational use rising from 20% of respondents in 2017 to 89% in 2026, and ChatGPT passed 1 billion weekly users in August 2026. Incident counts are rising too, but they record what is reported, not what happens, and the main public database could not be fully refreshed for this edition.

Key takeaways
  • Compute is the longest consistent series: about 25 orders of magnitude from 1950 to 2026 in Epoch AI’s largest-run-of-the-year data.
  • Growth sped up around 2010, from a 20-month doubling to roughly 6 months, and has run at about 5× a year for frontier language models since 2020.
  • A fixed capability gets cheaper fast; the newest frontier stays expensive to build and to buy.
  • Investment is cyclical even in this boom: corporate AI investment fell in 2022 and 2023 before more than doubling in 2025.
  • Compare like with like: weekly and monthly users, cloud-rental and amortised costs, and current and constant dollars are kept separate here.
Decade by decade

Every decade since 1950 raised the record for the largest training run except one: the 1970s, the decade of the first AI winter

The scoreboard below is generated from the dataset: for each decade it shows the eras it spanned, the largest training run in Epoch AI’s notable-models data, how that record compares with the previous decade’s, and the landmark events in our milestone register.

Exhibit 1
The largest training run of each year grew about 25 orders of magnitude since 1950
Training compute of the largest notable model each year, 1950–2026 (FLOP, log scale)
10010410810121016102010241028FLOP (log scale)1960197019801990200020102020Largest training run of the year: 40 FLOPLargest training run of the year: 6.9×10⁵ FLOPLargest training run of the year: 6×10⁸ FLOPLargest training run of the year: 7.2×10⁸ FLOPLargest training run of the year: 1.6×10⁶ FLOPLargest training run of the year: 2.2×10⁷ FLOPLargest training run of the year: 1.1×10⁶ FLOPLargest training run of the year: 1.1×10⁸ FLOPLargest training run of the year: 5.2×10⁶ FLOPLargest training run of the year: 2.7×10⁸ FLOPLargest training run of the year: 3.2×10⁸ FLOPLargest training run of the year: 6.7×10⁸ FLOPLargest training run of the year: 2.8×10¹⁰ FLOPLargest training run of the year: 3×10⁸ FLOPLargest training run of the year: 1.5×10¹² FLOPLargest training run of the year: 7.9×10¹⁰ FLOPLargest training run of the year: 7.5×10¹⁰ FLOPLargest training run of the year: 1.8×10¹³ FLOPLargest training run of the year: 1.3×10¹³ FLOPLargest training run of the year: 1.9×10¹³ FLOPLargest training run of the year: 2×10¹¹ FLOPLargest training run of the year: 8.8×10¹¹ FLOPLargest training run of the year: 3.2×10¹³ FLOPLargest training run of the year: 2.8×10¹² FLOPLargest training run of the year: 8×10¹² FLOPLargest training run of the year: 6.3×10¹⁵ FLOPLargest training run of the year: 6.3×10¹³ FLOPLargest training run of the year: 1.7×10¹⁵ FLOPLargest training run of the year: 2.8×10¹⁵ FLOPLargest training run of the year: 1.2×10¹⁴ FLOPLargest training run of the year: 7.5×10¹⁴ FLOPLargest training run of the year: 1.4×10¹⁸ FLOPLargest training run of the year: 1.6×10⁹ FLOPLargest training run of the year: 4.2×10¹⁵ FLOPLargest training run of the year: 5.4×10¹⁶ FLOPLargest training run of the year: 5.2×10¹⁶ FLOPLargest training run of the year: 6×10¹⁷ FLOPLargest training run of the year: 2.6×10¹⁸ FLOPLargest training run of the year: 3×10²⁰ FLOPLargest training run of the year: 3.8×10²⁰ FLOPLargest training run of the year: 6.6×10²¹ FLOPLargest training run of the year: 8.4×10²⁰ FLOPLargest training run of the year: 8.7×10²¹ FLOPLargest training run of the year: 1.1×10²³ FLOPLargest training run of the year: 3.1×10²³ FLOPLargest training run of the year: 2×10²⁴ FLOPLargest training run of the year: 2.7×10²⁴ FLOPLargest training run of the year: 5×10²⁵ FLOPLargest training run of the year: 3.8×10²⁵ FLOPLargest training run of the year: 5×10²⁶ FLOPLargest training run of the year: 9.4×10²⁶ FLOP
View the data
Largest training run of the year
195040
19576.9×10⁵
19596×10⁸
19607.2×10⁸
19621.6×10⁶
19632.2×10⁷
19651.1×10⁶
19661.1×10⁸
19755.2×10⁶
19802.7×10⁸
19833.2×10⁸
19866.7×10⁸
19872.8×10¹⁰
19883×10⁸
19891.5×10¹²
19907.9×10¹⁰
19917.5×10¹⁰
19921.8×10¹³
19931.3×10¹³
19941.9×10¹³
19952×10¹¹
19968.8×10¹¹
19973.2×10¹³
19982.8×10¹²
19998×10¹²
20006.3×10¹⁵
20016.3×10¹³
20031.7×10¹⁵
20042.8×10¹⁵
20051.2×10¹⁴
20067.5×10¹⁴
20071.4×10¹⁸
20081.6×10⁹
20094.2×10¹⁵
20105.4×10¹⁶
20115.2×10¹⁶
20126×10¹⁷
20132.6×10¹⁸
20143×10²⁰
20153.8×10²⁰
20166.6×10²¹
20178.4×10²⁰
20188.7×10²¹
20191.1×10²³
20203.1×10²³
20212×10²⁴
20222.7×10²⁴
20235×10²⁵
20243.8×10²⁵
20255×10²⁶
20269.4×10²⁶
Source: Epoch AI, Notable AI Models (downloaded 12 Nov 2026); Report AI selection of the annual maximum HIGH
Notes: Each point is the year’s largest estimated run among Epoch’s notable models; years with no notable model are skipped. Estimates carry Epoch confidence ratings; the 2026 value is rated ‘Likely’.REPORTAI report-ai.org

Two readings stand out. The 1970s record (about 5×10⁶ FLOP) sits below the 1960s one, the first AI winter showing up in the data. And the decade multiples are uneven because a single unusual model can set a decade’s record; Epoch’s fitted growth rates, in the compute section below, are the better guide to the trend. The 2020s figure covers 2020 to November 2026 only.

The decade scoreboard

DecadeEra(s)Largest estimated training runChange on previous decadeLandmarksConfidence
1950sFoundations, Early optimism6×10⁸ FLOP (Pandemonium (morse), 1959)—Turing’s ‘Computing Machinery and Intelligence’; Dartmouth proposal coins ‘artificial intelligence’; Dartmouth Summer Research Project; Rosenblatt’s perceptron unveiledDERIVED
1960sEarly optimism7.2×10⁸ FLOP (Perceptron (1960), 1960)×1.2Shakey the robot (SRI); ELIZA; Mansfield Amendment and ‘Perceptrons’DERIVED
1970sEarly optimism, First AI winter5.2×10⁶ FLOP (Cognitron, 1975)lower (÷139)Lighthill report; DARPA cuts AI fundingDERIVED
1980sExpert-systems boom, Second AI winter1.5×10¹² FLOP (Zip CNN, 1989)×2.9×10⁵R1/XCON at Digital Equipment Corporation; Japan’s Fifth Generation project (ICOT); Backpropagation popularised; Lisp-machine market collapses; NETtalkDERIVED
1990sSecond AI winter, Statistical machine learning3.2×10¹³ FLOP (LSTM, 1997)×21.1Deep Blue beats Kasparov; LSTM; LeNet-5DERIVED
2000sStatistical machine learning1.4×10¹⁸ FLOP (SB-LM, 2007)×4.6×10⁴DARPA Grand Challenge won; Deep belief netsDERIVED
2010sStatistical machine learning, Deep learning, Transformers and scale1.1×10²³ FLOP (AlphaStar, 2019)×7.4×10⁴First ImageNet challenge (ILSVRC); IBM Watson wins Jeopardy!; Siri ships on iPhone 4S; AlexNet wins ImageNet (+10 more)DERIVED
2020sTransformers and scale, Generative AI goes mainstream, Reasoning models and agents9.4×10²⁶ FLOP (GPT-6 Astra, 2026)×8,725GPT-3; AlphaFold 2 at CASP14; PaLM; Stable Diffusion released with open weights (+33 more)DERIVED

Generated from the AI Development Index dataset (Report AI calculation). Model counts refer to models in our compute series, not all models built.

1. Compute

Training compute: from 10⁶ FLOP for the Perceptron to about 10²⁷ for today’s frontier

Training compute is the number of arithmetic operations used to train a model. It is the measure most closely linked to capability gains in the deep-learning era and the one with the longest consistent record. All values are Epoch AI estimates; labs rarely disclose them, and Epoch rates each estimate’s confidence.

Growth rates: Epoch estimates a doubling roughly every 20 months before 2010 and about every 6 months afterwards (about 4–5× a year), and about 5× a year for frontier language models since 2020 (90% interval 4–6×).

Landmark models: estimated training compute

DateMeasureValueSourceConfidence
1950-07-02Theseus (Shannon’s maze mouse)40 FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
1957-01-01Perceptron Mark I6.9×10⁵ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
1960-03-30Perceptron (1960)7.2×10⁸ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
1987-06-06NETtalk2.8×10¹⁰ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
1998-11-01LeNet-52.8×10¹² FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2012-09-30AlexNet4.7×10¹⁷ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2014-09-10Seq2Seq LSTM5.6×10¹⁹ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2017-06-12Transformer7.4×10¹⁸ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2017-10-18AlphaGo Zero6.5×10²⁰ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2018-10-11BERT-Large2.9×10²⁰ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2019-02-14GPT-2 (1.5B)1.9×10²¹ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2020-05-28GPT-3 175B3.1×10²³ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2021-12-08Gopher 280B6.3×10²³ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2022-03-29Chinchilla 70B5.8×10²³ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2022-04-04PaLM 540B2.5×10²⁴ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2023-02-24LLaMA-65B5.5×10²³ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2023-03-15GPT-42.1×10²⁵ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2023-12-06Gemini 1.0 Ultra5×10²⁵ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2024-06-20Claude 3.5 Sonnet2.7×10²⁵ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2024-07-23Llama 3.1 405B3.8×10²⁵ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2024-08-13Grok-23×10²⁵ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2024-12-24DeepSeek-V33.3×10²⁴ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2025-02-17Grok 33.5×10²⁶ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2025-02-27GPT-4.53.8×10²⁶ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2025-07-09Grok 45×10²⁶ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2025-07-11Kimi K23×10²⁴ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2025-08-07GPT-56.6×10²⁵ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2026-04-24DeepSeek-V4-Pro9.7×10²⁴ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2026-09-03GPT-6 Astra9.4×10²⁶ FLOPEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH

Growth rates

DateMeasureValueSourceConfidence
pre-2010Growth rate: pre-2010 (doubling time)20 months per doublingSevilla et al. 2022, Compute Trends Across Three ErasMEDIUM
2012–2018Growth rate: OpenAI 2012–2018 (doubling time)3.4 months per doublingOpenAI, AI and Compute (May 2018)MEDIUM
2010–2022Growth rate: deep-learning era (doubling time)6 months per doublingSevilla et al. 2022MEDIUM
2010–May 2024Growth rate: notable models 2010–May 20244.1 × per yearEpoch AIMEDIUM
2010–2024Doubling time implied by 4.1×/yr5.9 months per doublingReport AI calculationDERIVED
2020–Feb 2026Growth rate: frontier language models since 20205 × per yearEpoch AI, Trends in AI (page ‘Updated Feb. 5, 2026’), opened 12 Nov 2026HIGH

Disclosed parameter counts

DateMeasureValueSourceConfidence
1957-01-01Perceptron Mark I1,000 parametersEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
1998-11-01LeNet-560,000 parametersEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2012-09-30AlexNet60M parametersEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2014-09-10Seq2Seq LSTM1.92B parametersEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2017-06-12Transformer (big)213M parametersEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2018-10-11BERT-Large340M parametersEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2019-02-14GPT-21.5B parametersEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2020-05-28GPT-3174.6B parametersEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2021-12-08Gopher280B parametersEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2022-03-29Chinchilla70B parametersEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2022-04-04PaLM540.35B parametersEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2023-02-24LLaMA (largest)65.2B parametersEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2024-07-23Llama 3.1 405B405B parametersEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2024-12-24DeepSeek-V3 (total; 37B active)671B parametersEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2025-07-11Kimi K2 (total)1T parametersEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH
2026-04-24DeepSeek V4-Pro (total)1.6T parametersEpoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)HIGH

GPT-4’s parameter count has never been disclosed; the widely quoted 1.76–1.8 trillion is a rumour.

2. Capability

Capability: benchmarks rise from near zero to saturation in one to five years

Benchmark scores are the most quoted and least stable measure of AI progress. Each test measures something narrow, results depend on settings, and most scores are reported by the developer. The table lists the dated results in our dataset; the companion report charts four of them and estimates how long each test lasted.

Benchmark results over time

DateMeasureValueSourceConfidence
2010ImageNet top-5 error — 2010 winner (NEC-UIUC)28.2 % errorILSVRC results (via search summaries)MEDIUM
2011ImageNet top-5 error — 2011 winner (XRCE)25.8 % errorILSVRC results (via search summaries)MEDIUM
Sep 2012ImageNet top-5 error — 2012 AlexNet15.3 % errorKrizhevsky et al. 2012; ILSVRC 2012MEDIUM
2013ImageNet top-5 error — 2013 winner (Clarifai)11.2 % errorILSVRC results (via search summaries)MEDIUM
2014ImageNet top-5 error — 2014 GoogLeNet6.67 % errorILSVRC results (via search summaries)MEDIUM
2 Sep 2014ImageNet human baseline (one annotator)5.1 % errorKarpathy blogMEDIUM
Dec 2015ImageNet top-5 error — 2015 ResNet3.57 % errorHe et al. 2015MEDIUM
2016ImageNet top-5 error — 2016 winner (Trimps-Soushen)2.99 % errorILSVRC results (via search summaries)MEDIUM
2017ImageNet top-5 error — 2017 SENet2.25 % errorHu et al. (SENet)MEDIUM
2010–2017ImageNet: error cut 2010→201712.5 × reductionReport AI calculationDERIVED
2020ARC-AGI-1 — 2020 Kaggle top20 % tasksARC Prize 2024 technical reportMEDIUM
2020MMLU — GPT-3 175B (few-shot)43.9 % correctHendrycks et al. MMLU leaderboardMEDIUM
Dec 2021MMLU — Gopher 280B60 % correctMMLU leaderboardMEDIUM
Mar 2022MMLU — Chinchilla 70B67.5 % correctMMLU leaderboardMEDIUM
2023SWE-bench (original) — best 20234.4 % resolvedStanford HAI AI Index 2025MEDIUM
Mar 2023MMLU — GPT-486.4 % correctOpenAI GPT-4 technical report (via aggregator)MEDIUM
Nov 2023GPQA — GPT-4 baseline (main set)39 % correctRein et al. 2023MEDIUM
2024ARC-AGI-1 — GPT-4o5 % tasks (semi-private)ARC Prize public leaderboardMEDIUM
Jun 2024GPQA Diamond — Claude 3.5 Sonnet59.4 % correctGPQA leaderboard snapshot (via search)MEDIUM
mid-2024ARC-AGI-1 — top score mid-202433 % tasksARC Prize 2024 technical reportMEDIUM
Jul 2024IMO — AlphaProof + AlphaGeometry 228 points of 42Google DeepMindMEDIUM
Sep 2024MMLU — o192.3 % correctOpenAI (o1 announcement)MEDIUM
Oct 2024SWE-bench Verified — Claude 3.5 Sonnet (best GA)49 % resolvedAnthropic, as cited on Report AIMEDIUM
Nov 2024ARC-AGI-1 — MindsAI (ARC Prize 2024)55.5 % tasks (private)ARC Prize 2024 technical reportMEDIUM
Nov 2024FrontierMath — best model at launch2 % solvedEpoch AI launch, as reportedMEDIUM
Dec 2024GPQA Diamond — o3 (preview)87.7 % correctOpenAIMEDIUM
Dec 2024SWE-bench Verified — o3 preview71.7 % resolvedOpenAI via Stanford HAI AI Index 2025MEDIUM
20 Dec 2024ARC-AGI-1 — o3 preview ($10k limit)75.7 % tasks (semi-private)ARC Prize FoundationMEDIUM
20 Dec 2024FrontierMath — o3 (OpenAI-reported)25.2 % solvedOpenAI, as reportedMEDIUM
Jan 2025Humanity’s Last Exam — launch (o1 / R1)9 % correctCAIS / Scale AIMEDIUM
Apr 2025FrontierMath — Epoch independent run of o310 % solvedEpoch AI via TechCrunchMEDIUM
Jul 2025IMO — Gemini Deep Think (certified gold)35 points of 42Google DeepMind / IMOMEDIUM
24 Nov 2025SWE-bench Verified — Claude Opus 4.580.9 % resolvedAnthropic (vendor)MEDIUM
Dec 2025GPQA Diamond — Gemini 3 Deep Think93.8 % correctGoogle (vendor)MEDIUM
Dec 2025Humanity’s Last Exam — Gemini 3 Deep Think (no tools)41 % correctR&D World tabulationMEDIUM
5 Dec 2025ARC-AGI-2 — Poetiq on Gemini 3 Pro54 % tasksARC Prize verification via BD Tech TalksMEDIUM
Feb 2026GPQA Diamond — Gemini 3.1 Pro94.3 % correctGoogle DeepMind model cardMEDIUM
Feb 2026ARC-AGI-2 — Gemini 3 Deep Think84.6 % tasksARC Prize verification via OfficeChaiMEDIUM
2026FrontierMath tiers 1–3 — Claude Fable 5 (Epoch)87 % solvedEpoch AI, as reported by The DecoderMEDIUM
Jul 2026IMO — multiple systems (perfect)42 points of 42Trade pressMEDIUM
1 Sep 2026SWE-bench Verified — Claude Opus 5 (Vals.ai)97 % resolvedVals.ai leaderboardMEDIUM
17 Oct 2026Humanity’s Last Exam — GPT-6 Astra54.8 % correctScale Labs leaderboard via BenchLM mirrorMEDIUM

Vendor-reported unless the source says otherwise. SWE-bench (2023) and SWE-bench Verified (2024 onward) are different tests.

3. Cost

Cost: buying a fixed capability gets cheaper every year, while building the frontier gets dearer

Two cost measures run in opposite directions. The price of using a model at a fixed level of capability has fallen by one to two orders of magnitude every few years; the cost of training the newest frontier model has risen by about 2.4–3.5× a year, according to Epoch AI.

Exhibit 2
GPT-4-level output fell about 187-fold in price in three and a half years
Cheapest blended price for a model at or above GPT-4’s original level, Mar 2023–Oct 2026 (US$ per million tokens, log scale)
10-1100101102US$ per M tokens (log scale)202420252026(GPT-4 launch): 37.5 US$ per M tokens(GPT-4 launch)(GPT-4 Turbo): 15 US$ per M tokens(GPT-4 Turbo)(DeepSeek-V3): 0.477 US$ per M tokens(DeepSeek-V3)(DeepSeek-V3.2-Exp): 0.315 US$ per M tokens(DeepSeek-V3.2-Exp)(GPT-6 Luna / Haiku 5.5): 0.2 US$ per M tokens(GPT-6 Luna / Haiku 5.5)
View the data
DateItemUS$ per M tokens
Mar 2023(GPT-4 launch)37.5
Nov 2023(GPT-4 Turbo)15
Dec 2024(DeepSeek-V3)0.477
Sep 2025(DeepSeek-V3.2-Exp)0.315
Oct 2026(GPT-6 Luna / Haiku 5.5)0.2
Source: provider list prices; Report AI token price index DERIVED
Notes: Blended at 3:1 input:output (Report AI method).REPORTAI report-ai.org

Inference prices

DateMeasureValueSourceConfidence
Nov 2022GPT-3.5-level (MMLU 64.8) cheapest price$20 per million tokensStanford HAI AI Index 2025MEDIUM
Mar 2023GPT-4-level cheapest blended price (GPT-4 launch)$37.5 per million tokens (3:1 blend)OpenAI list price $30/$60; Report AI blendDERIVED
Nov 2023GPT-4-level cheapest blended price (GPT-4 Turbo)$15 per million tokens (3:1 blend)OpenAI DevDay pricing; Report AI blendDERIVED
Nov 2022–Oct 2024GPT-3.5-level decline Nov 2022→Oct 2024286 × cheaperReport AI calculationDERIVED
Oct 2024GPT-3.5-level (MMLU 64.8) cheapest price$0.07 per million tokensStanford HAI AI Index 2025MEDIUM
Nov 2021–Nov 2024Price decline at fixed capability (a16z)10 × per yeara16z, LLMflationMEDIUM
Dec 2024GPT-4-level cheapest blended price (DeepSeek-V3)$0.477 per million tokens (3:1 blend)DeepSeek API; Report AI blendDERIVED
to Mar 2025Price decline at fixed capability (Epoch median)50 × per yearEpoch AI, LLM inference price trendsMEDIUM
Sep 2025GPT-4-level cheapest blended price (DeepSeek-V3.2-Exp)$0.315 per million tokens (3:1 blend)DeepSeek API; Report AI blendDERIVED
Mar 2023–Oct 2026GPT-4-level decline Mar 2023→Oct 2026188 × cheaperReport AI calculationDERIVED
three years to 2026Price decline at fixed capability (Epoch 2026)12.7 × per yearEpoch AI, The plunging price of thoughtDERIVED
1 Sep 2026Usage-weighted market price (Silicon Data index)$0.97 per million tokensSilicon Data via CNBCMEDIUM
Oct 2026GPT-4-level cheapest blended price (GPT-6 Luna / Haiku 5.5)$0.2 per million tokens (3:1 blend)Vendor prices as reported; Report AI blendDERIVED

Training costs

DateMeasureValueSourceConfidence
2012-09-30AlexNet — amortised cost (Epoch, 2023 USD)16 USD (2023 dollars)Epoch AI Notable AI Models CSV — ‘Training compute cost (2023 USD)’ (amortised hardware + HIGH
2017Transformer$930Stanford HAI AI Index 2024 (Epoch AI estimates, cloud-rental basis)MEDIUM
2017-06-12Transformer — amortised cost (Epoch, 2023 USD)438 USD (2023 dollars)Epoch AI Notable AI Models CSV — ‘Training compute cost (2023 USD)’ (amortised hardware + HIGH
2017-10-18AlphaGo Zero — amortised cost (Epoch, 2023 USD)1,214 USD (2023 dollars)Epoch AI Notable AI Models CSV — ‘Training compute cost (2023 USD)’ (amortised hardware + HIGH
2016–2024Growth rate: frontier training cost since 20162.4 × per yearEpoch AI (Cottier et al.), blog opened 12 Nov 2026HIGH
2020GPT-3 175B$4.3MStanford HAI AI Index 2024 (Epoch AI estimates, cloud-rental basis)MEDIUM
2020-05-28GPT-3 175B — amortised cost (Epoch, 2023 USD)2,116,866 USD (2023 dollars)Epoch AI Notable AI Models CSV — ‘Training compute cost (2023 USD)’ (amortised hardware + HIGH
2020-05-28GPT-3 175B — cloud-rental cost (Epoch)$4.5MEpoch AI Notable AI Models CSV — ‘Training compute cost (cloud)’HIGH
2022-04-04PaLM 540B — amortised cost (Epoch, 2023 USD)3,060,365 USD (2023 dollars)Epoch AI Notable AI Models CSV — ‘Training compute cost (2023 USD)’ (amortised hardware + HIGH
2022-04-04PaLM 540B — cloud-rental cost (Epoch)$14.5MEpoch AI Notable AI Models CSV — ‘Training compute cost (cloud)’HIGH
2020–Feb 2026Growth rate: frontier language-model training cost since 20203.5 × per yearEpoch AI, Trends in AI (updated 5 Feb 2026)HIGH
2023GPT-4$78.4MStanford HAI AI Index 2024 (Epoch AI estimates, cloud-rental basis)MEDIUM
2023-03-15GPT-4 — amortised cost (Epoch, 2023 USD)37,334,305 USD (2023 dollars)Epoch AI Notable AI Models CSV — ‘Training compute cost (2023 USD)’ (amortised hardware + HIGH
2023-03-15GPT-4 — cloud-rental cost (Epoch)$81.4MEpoch AI Notable AI Models CSV — ‘Training compute cost (cloud)’HIGH
2023Gemini 1.0 Ultra$191.4MStanford HAI AI Index 2024 (Epoch AI estimates, cloud-rental basis)MEDIUM
2023-12-06Gemini 1.0 Ultra — amortised cost (Epoch, 2023 USD)30,719,420 USD (2023 dollars)Epoch AI Notable AI Models CSV — ‘Training compute cost (2023 USD)’ (amortised hardware + HIGH
2023-12-06Gemini 1.0 Ultra — cloud-rental cost (Epoch)$206.1MEpoch AI Notable AI Models CSV — ‘Training compute cost (cloud)’HIGH
2024Llama 3.1 405B$170.0MStanford HAI AI Index 2025 (Epoch)MEDIUM
2024-07-23Llama 3.1 405B — amortised cost (Epoch, 2023 USD)52,885,434 USD (2023 dollars)Epoch AI Notable AI Models CSV — ‘Training compute cost (2023 USD)’ (amortised hardware + HIGH
2024-07-23Llama 3.1 405B — cloud-rental cost (Epoch)$176.0MEpoch AI Notable AI Models CSV — ‘Training compute cost (cloud)’HIGH
2024Grok-2$107.0MStanford HAI AI Index 2025 (Epoch), via secondary coverageMEDIUM
2024-08-13Grok-2 — amortised cost (Epoch, 2023 USD)31,602,311 USD (2023 dollars)Epoch AI Notable AI Models CSV — ‘Training compute cost (2023 USD)’ (amortised hardware + HIGH
Dec 2024DeepSeek-V3 (stated final-run GPU rental)$5.6MDeepSeek technical reportMEDIUM
2024-12-24DeepSeek-V3 — amortised cost (Epoch, 2023 USD)5,390,000 USD (2023 dollars)Epoch AI Notable AI Models CSV — ‘Training compute cost (2023 USD)’ (amortised hardware + HIGH
2017–2025Transformer → Grok 4 cost multiple885,000 × (multiple)Report AI calculationDERIVED
2025-02-17Grok 3 — amortised cost (Epoch, 2023 USD)217,835,546 USD (2023 dollars)Epoch AI Notable AI Models CSV — ‘Training compute cost (2023 USD)’ (amortised hardware + HIGH
2025-02-17Grok 3 — cloud-rental cost (Epoch)$849.7MEpoch AI Notable AI Models CSV — ‘Training compute cost (cloud)’HIGH
2025-02-27GPT-4.5 — amortised cost (Epoch, 2023 USD)366,010,628 USD (2023 dollars)Epoch AI Notable AI Models CSV — ‘Training compute cost (2023 USD)’ (amortised hardware + HIGH
mid-2025Largest training run, mid-2025 (AI Index 2026 claim)$120.0MStanford HAI AI Index 2026, via remio.ai (single secondary source)MEDIUM
2025Grok 4$490.0MEpoch AI data insight ‘What did it take to train Grok 4?’ (12 Sep 2025), opened 12 Nov 202HIGH
2025-07-09Grok 4 — amortised cost (Epoch, 2023 USD)387,842,678 USD (2023 dollars)Epoch AI Notable AI Models CSV — ‘Training compute cost (2023 USD)’ (amortised hardware + HIGH

Amortised (hardware depreciation plus energy, 2023 dollars) and cloud-rental estimates are different methods and are labelled separately.

4. Capital

Capital: corporate AI investment rose about 29-fold in real terms from 2013 to 2025

Investment figures come from the Stanford HAI AI Index (Quid data), republished by Our World in Data in constant 2021 dollars. Each AI Index edition restates earlier years, so current-dollar figures from different editions should not be joined into one series; the constant-dollar series below is consistent.

Exhibit 3
Private AI investment doubled in 2025
Global private investment in AI, 2013–2025 (US$ billion, constant 2021 dollars)
0100200300US$ B (2021)2013Global private AI investment, 2013: 6.01 US$ B (2021)6.012014Global private AI investment, 2014: 10.9 US$ B (2021)10.92015Global private AI investment, 2015: 15.3 US$ B (2021)15.32016Global private AI investment, 2016: 19.3 US$ B (2021)19.32017Global private AI investment, 2017: 28.4 US$ B (2021)28.42018Global private AI investment, 2018: 46.5 US$ B (2021)46.52019Global private AI investment, 2019: 61.7 US$ B (2021)61.72020Global private AI investment, 2020: 77.3 US$ B (2021)77.32021Global private AI investment, 2021: 145 US$ B (2021)1452022Global private AI investment, 2022: 105 US$ B (2021)1052023Global private AI investment, 2023: 92.8 US$ B (2021)92.82024Global private AI investment, 2024: 131 US$ B (2021)1312025Global private AI investment, 2025: 290 US$ B (2021)290
View the data
Global private AI investment
20136.01
201410.9
201515.3
201619.3
201728.4
201846.5
201961.7
202077.3
2021145
2022105
202392.8
2024131
2025290
Source: Our World in Data, from Stanford HAI AI Index (Quid data) HIGH
Notes: Private investment only. The AI Index’s current-dollar figure for 2025 is $344.7B.REPORTAI report-ai.org

Corporate, generative-AI and venture investment

DateMeasureValueSourceConfidence
2013Global corporate AI investment — Total (OWID, constant 2021 US$)16,947,474,940 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2014Global corporate AI investment — Total (OWID, constant 2021 US$)21,793,344,195 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2015Global corporate AI investment — Total (OWID, constant 2021 US$)29,072,882,368 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2016Global corporate AI investment — Total (OWID, constant 2021 US$)38,183,074,315 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2017Global corporate AI investment — Total (OWID, constant 2021 US$)59,385,234,310 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2018Global corporate AI investment (AI Index 2023 ed.)$79.6BStanford HAI AI Index 2023MEDIUM
2018Global corporate AI investment — Total (OWID, constant 2021 US$)85,918,080,126 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2019Generative-AI private investment$840.0MAI Index 2024 via ForklogMEDIUM
2019Global corporate AI investment — Total (OWID, constant 2021 US$)109,455,530,391 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2020Global private AI investment (AI Index 2022 ed.)$46.0BStanford HAI AI Index 2022 via IEEE SpectrumMEDIUM
2020Global corporate AI investment — Total (OWID, constant 2021 US$)232,293,505,123 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2021Global corporate AI investment (AI Index 2023 ed.)$276.1BStanford HAI AI Index 2023MEDIUM
2021Global private AI investment (AI Index 2022 ed.)$93.5BStanford HAI AI Index 2022MEDIUM
2021Global private AI investment (AI Index 2023 ed.)$125.4BStanford HAI AI Index 2023MEDIUM
2021Global corporate AI investment — Total (OWID, constant 2021 US$)360,729,994,297 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2022Global corporate AI investment (AI Index 2023 ed.)$189.6BStanford HAI AI Index 2023 via IEEE SpectrumMEDIUM
2022Global private AI investment (AI Index 2023 ed.)$91.9BStanford HAI AI Index 2023MEDIUM
2022Generative-AI private investment$3.0BStanford HAI (AI Index 2024 summary)MEDIUM
2022Global corporate AI investment — Total (OWID, constant 2021 US$)234,484,815,268 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2023Generative-AI private investment$25.2BStanford HAI AI Index 2024MEDIUM
2023Global corporate AI investment — Total (OWID, constant 2021 US$)178,748,319,573 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2024Global corporate AI investment (AI Index 2025 ed.)$252.3BStanford HAI AI Index 2025MEDIUM
2024Global private AI investment (AI Index 2026 ed., implied 2024 base)$151.5BReport AI calculation from AI Index 2026 (+127.5% to $344.66B)DERIVED
2024Generative-AI private investment$33.9BStanford HAI AI Index 2025MEDIUM
2024Global AI venture funding (Crunchbase)$100.0BCrunchbase NewsMEDIUM
2024Global corporate AI investment — Total (OWID, constant 2021 US$)218,563,067,882 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2025Global corporate AI investment (AI Index 2026 ed.)$581.7BStanford HAI AI Index 2026MEDIUM
2025Global private AI investment (AI Index 2026 ed.)$344.7BStanford HAI AI Index 2026MEDIUM
2025Generative-AI private investment$170.9BAI Index 2026 via single secondary summaryMEDIUM
2025Global AI venture funding (Crunchbase)$202.3BCrunchbase NewsMEDIUM
2025Global corporate AI investment — Total (OWID, constant 2021 US$)489,591,464,709 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
Jan–Sep 2026Global AI venture funding (Crunchbase), Jan–Sep$488.0BCrunchbase quarterly reports; Report AI sumDERIVED

Private AI investment, United States and China

DateMeasureValueSourceConfidence
2013US private AI investment (OWID, constant 2021 US$)4,211,478,899 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2013China private AI investment (OWID, constant 2021 US$)717,196,154 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2014US private AI investment (OWID, constant 2021 US$)7,963,123,765 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2014China private AI investment (OWID, constant 2021 US$)771,392,259 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2015US private AI investment (OWID, constant 2021 US$)10,632,666,176 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2015China private AI investment (OWID, constant 2021 US$)2,385,249,652 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2016US private AI investment (OWID, constant 2021 US$)11,464,266,076 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2016China private AI investment (OWID, constant 2021 US$)5,102,962,753 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2017US private AI investment (OWID, constant 2021 US$)15,548,779,419 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2017China private AI investment (OWID, constant 2021 US$)7,314,146,387 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2018US private AI investment (OWID, constant 2021 US$)22,716,604,134 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2018China private AI investment (OWID, constant 2021 US$)15,069,709,833 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2019US private AI investment (OWID, constant 2021 US$)33,994,613,790 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2019China private AI investment (OWID, constant 2021 US$)14,895,814,015 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2020US private AI investment (OWID, constant 2021 US$)46,294,131,458 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2020China private AI investment (OWID, constant 2021 US$)17,871,570,528 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2021US private AI investment (OWID, constant 2021 US$)86,855,560,303 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2021China private AI investment (OWID, constant 2021 US$)25,891,134,262 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2022US private AI investment (OWID, constant 2021 US$)56,651,417,674 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2022China private AI investment (OWID, constant 2021 US$)13,951,917,588 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2023US private AI investment (OWID, constant 2021 US$)64,353,629,904 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2023China private AI investment (OWID, constant 2021 US$)8,419,883,132 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2024US private AI investment$109.1BStanford HAI AI Index 2025MEDIUM
2024US private AI investment (OWID, constant 2021 US$)94,629,349,595 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2024China private AI investment (OWID, constant 2021 US$)8,109,463,468 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2025US private AI investment$285.9BStanford HAI AI Index 2026MEDIUM
2025US private AI investment (OWID, constant 2021 US$)240,612,883,142 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
2025China private AI investment (OWID, constant 2021 US$)10,446,895,788 USD (constant 2021)Our World in Data, from Quid via Stanford HAI AI Index Report (2026); constant 2021 US$ (CHIGH
5. Adoption and research output

Adoption: from a specialist tool to a billion weekly users

Adoption is measured two ways here: by organisations (McKinsey’s annual survey) and by people (ChatGPT weekly active users disclosed by OpenAI). Neither measures depth of use or financial impact. Research output is shown through notable-model counts by affiliation and country, and AI publications and patents.

Organisational and consumer adoption

DateMeasureValueSourceConfidence
2017Organisations using AI in ≥1 function (McKinsey)20 % of respondentsMcKinsey Global Survey on AI (via AI Index 2024 / McKinsey 2024 chart)MEDIUM
2018Organisations using AI in ≥1 function (McKinsey)47 % of respondentsMcKinsey Global Survey on AI (via McKinsey 2024 chart)MEDIUM
2019Organisations using AI in ≥1 function (McKinsey)58 % of respondentsMcKinsey Global Survey on AI (via McKinsey 2024 chart)MEDIUM
2020Organisations using AI in ≥1 function (McKinsey)50 % of respondentsMcKinsey Global Survey on AI, The state of AI in 2020MEDIUM
2021Organisations using AI in ≥1 function (McKinsey)56 % of respondentsMcKinsey Global Survey on AI, The state of AI in 2021MEDIUM
2022Organisations using AI in ≥1 function (McKinsey)50 % of respondentsMcKinsey Global Survey on AI (via Stanford AI Index 2024)MEDIUM
2023Organisations using AI in ≥1 function (McKinsey)55 % of respondentsMcKinsey Global Survey on AI (via Stanford AI Index 2024)MEDIUM
Jan 2023ChatGPT monthly users (UBS estimate)100M users (monthly)UBS analyst note using Similarweb data, via ReutersMEDIUM
6 Nov 2023ChatGPT weekly active users100M users (weekly)OpenAI DevDayMEDIUM
Feb–Mar 2024 surveyOrganisations using AI in ≥1 function (McKinsey)72 % of respondentsMcKinsey Global Survey on AI, early 2024MEDIUM
Jul 2024 surveyOrganisations using AI in ≥1 function (McKinsey)78 % of respondentsMcKinsey Global Survey on AIMEDIUM
Aug 2024ChatGPT weekly active users200M users (weekly)OpenAI via AxiosMEDIUM
Dec 2024ChatGPT weekly active users300M users (weekly)OpenAIMEDIUM
Feb 2025ChatGPT weekly active users400M users (weekly)OpenAI via Reuters/TechCrunchMEDIUM
end Mar 2025ChatGPT weekly active users500M users (weekly)Business Insider via AOLMEDIUM
Jun–Jul 2025 surveyOrganisations using AI in ≥1 function (McKinsey)88 % of respondentsMcKinsey Global Survey on AI, State of AI 2025MEDIUM
Jul 2025ChatGPT weekly active users700M users (weekly)OpenAIMEDIUM
6 Oct 2025ChatGPT weekly active users800M users (weekly)OpenAI DevDay via TechCrunchMEDIUM
Nov 2023–Aug 2026ChatGPT weekly users: 100M → 1B2.83 yearsReport AI calculationDERIVED
27 Feb 2026ChatGPT weekly active users900M users (weekly)OpenAI via TechCrunchMEDIUM
May–Jun 2026 surveyOrganisations using AI in ≥1 function (McKinsey)89 % of respondentsMcKinsey Global Survey on AI, State of AI 2026MEDIUM
31 Aug 2026ChatGPT weekly active users1,000M users (weekly)OpenAIMEDIUM

Notable AI models per year, all and by country

DateMeasureValueSourceConfidence
2010Notable AI models — all (Epoch CSV count)10 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2010Notable AI models with a US-based organisation (Epoch CSV count)4 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2010Notable AI models with a China-based organisation (Epoch CSV count)0 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2011Notable AI models — all (Epoch CSV count)11 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2011Notable AI models with a US-based organisation (Epoch CSV count)5 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2011Notable AI models with a China-based organisation (Epoch CSV count)0 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2012Notable AI models — all (Epoch CSV count)13 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2012Notable AI models with a US-based organisation (Epoch CSV count)6 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2012Notable AI models with a China-based organisation (Epoch CSV count)0 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2013Notable AI models — all (Epoch CSV count)19 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2013Notable AI models with a US-based organisation (Epoch CSV count)13 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2013Notable AI models with a China-based organisation (Epoch CSV count)0 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2014Notable AI models — all (Epoch CSV count)33 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2014Notable AI models with a US-based organisation (Epoch CSV count)16 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2014Notable AI models with a China-based organisation (Epoch CSV count)3 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2015Notable AI models — all (Epoch CSV count)29 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2015Notable AI models with a US-based organisation (Epoch CSV count)19 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2015Notable AI models with a China-based organisation (Epoch CSV count)4 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2016Notable AI models — all (Epoch CSV count)49 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2016Notable AI models with a US-based organisation (Epoch CSV count)36 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2016Notable AI models with a China-based organisation (Epoch CSV count)6 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2017Notable AI models — all (Epoch CSV count)55 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2017Notable AI models with a US-based organisation (Epoch CSV count)29 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2017Notable AI models with a China-based organisation (Epoch CSV count)10 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2018Notable AI models — all (Epoch CSV count)37 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2018Notable AI models with a US-based organisation (Epoch CSV count)27 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2018Notable AI models with a China-based organisation (Epoch CSV count)1 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2019Notable AI models — all (Epoch CSV count)68 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2019Notable AI models with a US-based organisation (Epoch CSV count)56 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2019Notable AI models with a China-based organisation (Epoch CSV count)6 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2020Notable AI models — all (Epoch CSV count)50 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2020Notable AI models with a US-based organisation (Epoch CSV count)37 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2020Notable AI models with a China-based organisation (Epoch CSV count)8 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2021Notable AI models — all (Epoch CSV count)79 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2021Notable AI models with a US-based organisation (Epoch CSV count)54 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2021Notable AI models with a China-based organisation (Epoch CSV count)21 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2022Notable AI models — all (Epoch CSV count)91 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2022Notable AI models with a US-based organisation (Epoch CSV count)68 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2022Notable AI models with a China-based organisation (Epoch CSV count)12 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2023Notable AI models — all (Epoch CSV count)119 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2023Notable AI models with a US-based organisation (Epoch CSV count)86 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2023Notable AI models with a China-based organisation (Epoch CSV count)23 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2024Notable AI models — all (Epoch CSV count)99 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2024Notable AI models with a US-based organisation (Epoch CSV count)58 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2024Notable AI models with a China-based organisation (Epoch CSV count)31 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2025Notable AI models — all (Epoch CSV count)109 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2025Notable AI models with a US-based organisation (Epoch CSV count)65 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2025Notable AI models with a China-based organisation (Epoch CSV count)36 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2026 (to 12 Nov)Notable AI models — all (Epoch CSV count)95 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2026 (to 12 Nov)Notable AI models with a US-based organisation (Epoch CSV count)57 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED
2026 (to 12 Nov)Notable AI models with a China-based organisation (Epoch CSV count)31 modelsReport AI count of Epoch AI, Notable AI Models database (CSV downloaded 12 Nov 2026)DERIVED

Notable AI systems by affiliation (industry, academia, both)

DateMeasureValueSourceConfidence
2003Notable AI systems by affiliation — Academia (OWID/Epoch)4 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2003Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)0 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2003Notable AI systems by affiliation — Industry (OWID/Epoch)0 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2003Notable AI systems by affiliation — Other (OWID/Epoch)0 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2004Notable AI systems by affiliation — Academia (OWID/Epoch)1 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2004Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)0 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2004Notable AI systems by affiliation — Industry (OWID/Epoch)0 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2004Notable AI systems by affiliation — Other (OWID/Epoch)0 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2005Notable AI systems by affiliation — Academia (OWID/Epoch)1 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2005Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)0 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2005Notable AI systems by affiliation — Industry (OWID/Epoch)1 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2005Notable AI systems by affiliation — Other (OWID/Epoch)0 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2006Notable AI systems by affiliation — Academia (OWID/Epoch)8 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2006Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)1 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2006Notable AI systems by affiliation — Industry (OWID/Epoch)0 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2006Notable AI systems by affiliation — Other (OWID/Epoch)0 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2007Notable AI systems by affiliation — Academia (OWID/Epoch)3 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2007Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)0 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2007Notable AI systems by affiliation — Industry (OWID/Epoch)3 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2007Notable AI systems by affiliation — Other (OWID/Epoch)0 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2008Notable AI systems by affiliation — Academia (OWID/Epoch)4 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2008Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)0 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2008Notable AI systems by affiliation — Industry (OWID/Epoch)2 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2008Notable AI systems by affiliation — Other (OWID/Epoch)0 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2009Notable AI systems by affiliation — Academia (OWID/Epoch)5 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2009Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)0 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2009Notable AI systems by affiliation — Industry (OWID/Epoch)6 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2009Notable AI systems by affiliation — Other (OWID/Epoch)0 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2010Notable AI systems by affiliation — Academia (OWID/Epoch)8 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2010Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)1 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2010Notable AI systems by affiliation — Industry (OWID/Epoch)1 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2010Notable AI systems by affiliation — Other (OWID/Epoch)0 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2011Notable AI systems by affiliation — Academia (OWID/Epoch)8 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2011Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)3 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2011Notable AI systems by affiliation — Industry (OWID/Epoch)0 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2011Notable AI systems by affiliation — Other (OWID/Epoch)0 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2012Notable AI systems by affiliation — Academia (OWID/Epoch)8 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2012Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)2 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2012Notable AI systems by affiliation — Industry (OWID/Epoch)3 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2012Notable AI systems by affiliation — Other (OWID/Epoch)0 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2013Notable AI systems by affiliation — Academia (OWID/Epoch)10 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2013Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)2 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2013Notable AI systems by affiliation — Industry (OWID/Epoch)7 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2013Notable AI systems by affiliation — Other (OWID/Epoch)0 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2014Notable AI systems by affiliation — Academia (OWID/Epoch)21 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2014Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)6 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2014Notable AI systems by affiliation — Industry (OWID/Epoch)5 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2014Notable AI systems by affiliation — Other (OWID/Epoch)1 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2015Notable AI systems by affiliation — Academia (OWID/Epoch)9 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2015Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)5 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2015Notable AI systems by affiliation — Industry (OWID/Epoch)13 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2015Notable AI systems by affiliation — Other (OWID/Epoch)1 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2016Notable AI systems by affiliation — Academia (OWID/Epoch)18 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2016Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)8 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2016Notable AI systems by affiliation — Industry (OWID/Epoch)21 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2016Notable AI systems by affiliation — Other (OWID/Epoch)2 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2017Notable AI systems by affiliation — Academia (OWID/Epoch)18 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2017Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)10 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2017Notable AI systems by affiliation — Industry (OWID/Epoch)26 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2017Notable AI systems by affiliation — Other (OWID/Epoch)1 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2018Notable AI systems by affiliation — Academia (OWID/Epoch)4 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2018Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)8 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2018Notable AI systems by affiliation — Industry (OWID/Epoch)23 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2018Notable AI systems by affiliation — Other (OWID/Epoch)2 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2019Notable AI systems by affiliation — Academia (OWID/Epoch)12 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2019Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)21 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2019Notable AI systems by affiliation — Industry (OWID/Epoch)30 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2019Notable AI systems by affiliation — Other (OWID/Epoch)5 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2020Notable AI systems by affiliation — Academia (OWID/Epoch)9 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2020Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)12 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2020Notable AI systems by affiliation — Industry (OWID/Epoch)25 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2020Notable AI systems by affiliation — Other (OWID/Epoch)4 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2021Notable AI systems by affiliation — Academia (OWID/Epoch)9 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2021Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)20 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2021Notable AI systems by affiliation — Industry (OWID/Epoch)46 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2021Notable AI systems by affiliation — Other (OWID/Epoch)4 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2022Notable AI systems by affiliation — Academia (OWID/Epoch)17 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2022Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)19 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2022Notable AI systems by affiliation — Industry (OWID/Epoch)50 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2022Notable AI systems by affiliation — Other (OWID/Epoch)5 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2023Notable AI systems by affiliation — Academia (OWID/Epoch)18 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2023Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)29 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2023Notable AI systems by affiliation — Industry (OWID/Epoch)69 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2023Notable AI systems by affiliation — Other (OWID/Epoch)3 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2024Notable AI systems by affiliation — Academia (OWID/Epoch)3 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2024Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)12 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2024Notable AI systems by affiliation — Industry (OWID/Epoch)81 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2024Notable AI systems by affiliation — Other (OWID/Epoch)2 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2025Notable AI systems by affiliation — Academia (OWID/Epoch)2 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2025Notable AI systems by affiliation — Academia and industry collaboration (OWID/Epoch)5 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2025Notable AI systems by affiliation — Industry (OWID/Epoch)97 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH
2025Notable AI systems by affiliation — Other (OWID/Epoch)1 systems per yearOur World in Data, from Epoch AI (2026); 2026 data incomplete (Epoch update 4 Sep 2026)HIGH

AI publications and patents

DateMeasureValueSourceConfidence
2010AI patents (AI Index 2025 series)3,833 patentsStanford HAI AI Index 2025MEDIUM
2013AI publications (all venues)102,000 publicationsStanford HAI AI Index 2025MEDIUM
2022AI patents granted (AI Index 2024 series)62,264 patentsStanford HAI AI Index 2024MEDIUM
2023AI publications (all venues)242,000 publicationsStanford HAI AI Index 2025MEDIUM
2023AI patents (AI Index 2025 series)122,511 patentsStanford HAI AI Index 2025MEDIUM
2013–2023AI publications growth 2013→20232.37 × (multiple)Report AI calculationDERIVED
6. Risk and resources

Risk and resources: incidents and electricity use are rising, but are the least complete series

Reported incidents measure what is reported to the AI Incident Database, which grows with attention as well as harm. Electricity figures for data centres cover all workloads, not only AI; AI-optimised servers are estimated separately where a source provides it.

Reported AI incidents

DateMeasureValueSourceConfidence
2012AI incidents reported (AI Incident Database)10 incidents (upper bound)Stanford HAI AI Index 2026 (AIID data)MEDIUM
2023AI incidents reported (AI Incident Database)149 incidentsReport AI calculation from AI Index 2025 (+56.4% to 233)DERIVED
2024AI incidents reported (AI Incident Database)233 incidentsStanford HAI AI Index 2025MEDIUM
2025AI incidents reported (AI Incident Database)362 incidentsStanford HAI AI Index 2026MEDIUM
Aug–Oct 2026New AIID entries, Aug–Oct179 entriesReport AI count of AIID IDsDERIVED

Data-centre electricity

DateMeasureValueSourceConfidence
2023US data-centre electricity (LBNL)176 TWhLBNL (Dec 2024), as cited on Report AIMEDIUM
2024Global data-centre electricity (IEA)415 TWhIEA, Energy and AI (Apr 2025)MEDIUM
2025Global data-centre electricity (IEA)485 TWhIEA, Key Questions on Energy and AI (Apr 2026)MEDIUM
2025Global data-centre electricity (Gartner)447 TWhGartner (10 Jun 2026)MEDIUM
2025AI-optimised server electricity (Gartner)95 TWhGartner via Network WorldMEDIUM
2026 forecastGlobal data-centre electricity (Gartner forecast)565 TWhGartner (10 Jun 2026)MEDIUM
projection to 2030Global data-centre electricity (IEA projection)945 TWhIEA, Energy and AI (Apr 2025)MEDIUM
2024–2030IEA 2024→2030 growth0.147 CAGR (fraction)Report AI calculationDERIVED
projection to 2035Global data-centre electricity (IEA Base Case)1,200 TWhIEA, Energy and AIMEDIUM
Register

The milestone register: 70 dated events from 1943 to 2026

Every milestone below carries its date, what it showed, a source and a confidence rating. Events from 2022 onward are covered in more detail in the State of AI timeline and annual editions.

All milestones

DateMilestoneWhat it showedSourceConfidence
1943McCulloch–Pitts neuronFirst mathematical model of a network of simplified neurons computing logical functions.McCulloch & Pitts, Bulletin of Mathematical Biophysics 5MEDIUM
1950-10Turing’s ‘Computing Machinery and Intelligence’Replaces ‘Can machines think?’ with the imitation game (Turing test). Turing predicted that in about 50 years machines with ~10^9 storage would leave an average interrogator no more than a 70% chance of a correct identification after five minutes.Turing, Mind 59(236)MEDIUM
1955-08-31Dartmouth proposal coins ‘artificial intelligence’McCarthy, Minsky, Rochester and Shannon propose a summer study; the document introduces the term. Proposed as a ‘2 month, 10 man study’.AI Magazine reprint of the proposalMEDIUM
1956-06-18Dartmouth Summer Research ProjectConventional founding event of AI as a field. Ran roughly 18 Jun–17 Aug 1956 (about eight weeks) per Solomonoff’s notes; other accounts say six weeks.Wikipedia, Dartmouth workshop (citing Solomonoff)MEDIUM
1958-07Rosenblatt’s perceptron unveiledFirst learning machine to attract mass publicity; the US Navy’s ONR demonstrated it in July 1958 (a simulation on an IBM 704). Mark I Perceptron used a 400-sensor ‘retina’ and ~1,000 weights; Epoch estimates its training compute at ~6.95×10^5 FLOP.Rosenblatt, Psychological Review 65(6) 1958; Wikipedia Mark I PerceptronMEDIUM
1966Shakey the robot (SRI)First mobile robot to reason about its own actions; its STRIPS planner became the basis for later automated planners.Holloway, Making Things ThinkMEDIUM
1966-01ELIZAWeizenbaum’s pattern-matching ‘therapist’ program; first widely known chatbot and first demonstration of people over-attributing understanding to software.Weizenbaum, Communications of the ACM 9(1), 1966MEDIUM
1969Mansfield Amendment and ‘Perceptrons’US law pushes DARPA toward mission-oriented research; Minsky & Papert’s book highlights single-layer perceptron limits.Wikipedia, AI winterMEDIUM
1973Lighthill reportUK Science Research Council review concludes AI had failed its grand promises; UK funding is sharply curtailed. Debated on BBC TV in June 1973.Holloway / Wikipedia AI winterMEDIUM
1974DARPA cuts AI fundingMarks the start of the first AI winter in the US. DARPA cancelled ~$3M a year of Speech Understanding Research grants at Carnegie Mellon.Wikipedia, AI winterMEDIUM
1980R1/XCON at Digital Equipment CorporationFirst commercially significant expert system; starts the expert-systems boom. Most sources put savings at ~$25M a year; by 1986 it had processed 80,000 orders at 95–98% accuracy.McDermott (AAAI 1980); secondary historiesMEDIUM
1982Japan’s Fifth Generation project (ICOT)Ten-year MITI programme for logic-programming computers; spurred Western counter-programmes. Initial budget ~¥54 billion over ten years; total spending a little under ¥57 billion (~$320M) by the end.Wikipedia, Fifth Generation Computer SystemsMEDIUM
1986-10-09Backpropagation popularisedRumelhart, Hinton & Williams show multi-layer networks learn useful internal representations. Nature 323:533–536.NatureMEDIUM
1987Lisp-machine market collapses; NETtalkEnd of dedicated AI hardware boom marks the second AI winter; same year NETtalk shows a neural net learning to pronounce text. NETtalk training compute ~2.8×10^10 FLOP (Epoch).Holloway; Epoch AI via searchMEDIUM
1997-05-11Deep Blue beats KasparovFirst defeat of a reigning world chess champion by a computer in a match under tournament conditions. Won the rematch 3½–2½ (two wins, one loss, three draws); evaluated ~200 million positions per second. Kasparov had won the 1996 match 4–2.IBM / History.com, via searchMEDIUM
1997-11LSTMHochreiter & Schmidhuber’s long short-term memory, the recurrent architecture behind most sequence models until the Transformer.Neural Computation 9(8)MEDIUM
1998-11LeNet-5LeCun et al.’s convolutional network for handwriting recognition, deployed for cheque reading. Training compute ~2.8×10^12 FLOP (Epoch, ‘Confident’).Epoch AI via search summaryMEDIUM
2005-10DARPA Grand Challenge wonStanford’s Stanley completes a 132-mile desert course autonomously, a year after no vehicle finished. 6 h 53 min 58 s; in 2004 the best vehicle managed ~7.4 miles.DARPA / Stanford Racing TeamMEDIUM
2006-07Deep belief netsHinton, Osindero & Teh show deep networks can be trained layer by layer, reviving neural-network research.Neural Computation 18(7):1527–1554MEDIUM
2010First ImageNet challenge (ILSVRC)Establishes the benchmark that later showcased deep learning. Winning top-5 error 28.2% (NEC-UIUC).ILSVRC results via searchMEDIUM
2011-02-16IBM Watson wins Jeopardy!Open-domain question answering beats the show’s two best champions on TV. Two-game totals: Watson $77,147; Ken Jennings $24,000; Brad Rutter $21,600.IBMMEDIUM
2011-10Siri ships on iPhone 4SFirst voice assistant in mainstream consumer use.Search summaryMEDIUM
2012-09-30AlexNet wins ImageNetGPU-trained deep convolutional net wins by a wide margin, triggering the deep-learning era. Top-5 error 15.3% vs 26.2% for the second-best entry; ~60M parameters; trained on two GTX 580 GPUs; ~4.7×10^17 FLOP (Epoch).Krizhevsky, Sutskever & Hinton, NIPS 2012; Epoch AIMEDIUM
2014-01-26Google acquires DeepMindBig Tech begins buying frontier AI labs. Price undisclosed; reported at $400M (Re/code) to more than $500M (The Information).TechCrunchMEDIUM
2014-06-10Generative adversarial networksGoodfellow et al. introduce GANs, the main image-generation method until diffusion models.arXiv 1406.2661MEDIUM
2014-09Sequence-to-sequence learningSutskever, Vinyals & Le show end-to-end neural machine translation with LSTMs. Training compute ~5.6×10^19 FLOP (Epoch).Epoch AI via searchMEDIUM
2015-12ResNet beats human-level ImageNet errorResidual networks win ILSVRC 2015. Top-5 error 3.57%, below the 5.1% single-annotator human estimate.He et al., arXiv 1512.03385MEDIUM
2016-03-15AlphaGo defeats Lee SedolFirst defeat of a 9-dan Go professional without handicap, a decade earlier than many expected. 4–1 in Seoul, 9–15 Mar 2016; AlphaGo had beaten Fan Hui 5–0 in Oct 2015 (Nature, 27 Jan 2016).Nature / pressMEDIUM
2017-06-12The Transformer‘Attention Is All You Need’ introduces the architecture behind every frontier language model since. 28.4 BLEU on WMT14 English–German; ~7.4×10^18 FLOP training compute (Epoch).arXiv 1706.03762MEDIUM
2017-10AlphaGo ZeroLearns Go from self-play alone, without human games; among the first models trained on >10^23 FLOP. ~6.5×10^20 FLOP of direct training compute (Epoch, ‘Confident’); OpenAI’s 2018 figure of ~3.4×10^23 included self-play data generation.Epoch AI Notable AI Models CSVMEDIUM
2018-05OpenAI ‘AI and Compute’First widely cited measurement of exploding training compute. More than 300,000× growth from AlexNet to AlphaGo Zero; 3.4-month doubling time (2012–2018).OpenAIMEDIUM
2018-10BERTPre-trained bidirectional Transformer resets most language-understanding benchmarks. BERT-Large ~340M parameters; ~2.8×10^20 FLOP (Epoch).Google; Epoch AIMEDIUM
2019-02-14GPT-2 and staged releaseOpenAI withholds the full model at first over misuse concerns, starting the release-norms debate. 1.5B parameters; ~1.9×10^21 FLOP (Epoch).OpenAI; Epoch AIMEDIUM
2020-05-28GPT-3Shows few-shot learning emerges from scale. 175B parameters; 3.14×10^23 FLOP; 43.9% on MMLU (few-shot); ~$4.3M estimated training cost (AI Index 2024).Brown et al., arXiv 2005.14165MEDIUM
2020-11-30AlphaFold 2 at CASP14Effectively solves single-chain protein structure prediction, a 50-year-old problem. Median 92.4 GDT across all targets (≈90 is considered competitive with experiment).Google DeepMindMEDIUM
2022-04PaLMLargest dense model of its time; strong reasoning results. 540B parameters; ~2.5×10^24 FLOP (Epoch).Google; Epoch AIMEDIUM
2022-08-22Stable Diffusion released with open weightsHigh-quality image generation runs on consumer hardware.Stability AI via secondary timelinesMEDIUM
2022-11-30ChatGPT launchesGenerative AI becomes a mass consumer product. 1M users in five days (OpenAI); UBS estimated 100M monthly users by January 2023 from Similarweb data.OpenAI; UBS via ReutersMEDIUM
2023-02-24Meta releases LLaMACapable model weights spread beyond big labs (leaked publicly in early March), seeding the open-weight ecosystem. 7B–65B parameters.Meta; secondary timelinesMEDIUM
2023-03-14GPT-4Frontier model with strong exam performance; sets the ‘GPT-4 class’ bar. 86.4% MMLU; ~2.1×10^25 FLOP and ~$78M training compute cost (Epoch/AI Index); OpenAI’s ‘top 10% of bar takers’ claim was later disputed.OpenAI technical reportMEDIUM
2023-03-22FLI ‘pause’ letterOpen letter calls for a six-month pause on training systems more powerful than GPT-4.Future of Life InstituteMEDIUM
2023-11-01Bletchley Park AI Safety SummitFirst intergovernmental summit on frontier-AI risk; Bletchley Declaration.UK GovernmentMEDIUM
2023-11-06ChatGPT reaches 100M weekly usersOpenAI’s first own disclosure of the 100M mark (weekly users). 100M weekly active users at OpenAI’s first DevDay.OpenAI DevDay via TechCrunchMEDIUM
2023-12Gemini 1.0 UltraGoogle’s first frontier multimodal model; most expensive training run estimated at the time. ~5.0×10^25 FLOP (Epoch, ‘Speculative’); ~$191M cloud-basis cost in AI Index 2024 (Epoch’s current CSV: $206M cloud, $30.7M amortised in 2023 USD).Epoch AI; Stanford AI Index 2024MEDIUM
2024-07Llama 3.1 405BOpen-weight model at near-frontier scale. ~3.8×10^25 FLOP (6 × 405B params × 15.6T tokens); ~$170M estimated cost (AI Index 2025).Epoch AI; Stanford AI Index 2025MEDIUM
2024-08-01EU AI Act enters into forceFirst comprehensive AI statute in a major economy; obligations phase in from Feb 2025. Maximum fines €35M or 7% of worldwide turnover.European CommissionMEDIUM
2024-09-12OpenAI o1-preview: reasoning modelsModels trained to ‘think’ before answering; inference-time compute becomes a scaling axis. OpenAI reported o1 solved 83% of AIME (the US IMO qualifier) problems with 64-sample consensus vs 13% for GPT-4o; o1-preview ~57%.OpenAIMEDIUM
2024-10-08Nobel Prizes for AI researchPhysics to Hopfield and Hinton (8 Oct); half of Chemistry to Hassabis and Jumper for AlphaFold (9 Oct).Nobel announcements via pressMEDIUM
2024-10-22Computer use (Anthropic)First frontier model offered with the ability to operate a desktop GUI; start of the agent era.AnthropicMEDIUM
2024-12-20o3 announcedLarge jumps on hard reasoning benchmarks. ARC-AGI-1 75.7% within the $10k limit (87.5% high-compute); FrontierMath 25.2% (OpenAI-run; Epoch’s own later run ~10%); SWE-bench Verified 71.7%.ARC Prize; OpenAI via pressMEDIUM
2025-01Stargate announcedPrivate US AI-infrastructure commitment marks the gigawatt-scale data-centre era. $500B over four years, $100B immediately.OpenAIMEDIUM
2025-01-20DeepSeek-R1Chinese open-weight reasoning model near the frontier at far lower prices. Nvidia lost ~$589B of market value on 27 Jan 2025 (Bloomberg; Reuters says ~$593B), a record one-day loss.Bloomberg; DeepSeekMEDIUM
2025-02Grok 3: first model above 10^26 FLOPEpoch’s first estimate of a training run above 10^26 FLOP. ~3.5×10^26 FLOP (Epoch CSV, ‘Likely’; ~80,000 H100s for ~3 months).Epoch AIHIGH
2025-07AI wins IMO gold (certified)Natural-language maths at olympiad gold level within contest time. 35/42 (5 of 6 problems), Gemini Deep Think, certified by the IMO.Google DeepMindMEDIUM
2025-07Grok 4: largest known 2025 training runLargest estimated run of 2025. ~5×10^26 FLOP and ~$490M (Epoch median: 246M H100-hours).Epoch AIHIGH
2025-08-07GPT-5Flagship launched at lower price than predecessors; Epoch estimates it used less training compute than GPT-4.5. ~6.6×10^25 FLOP (Epoch CSV, ‘Speculative’, CI 2e25–2e26) vs ~3.8×10^26 for GPT-4.5; $1.25 / $10 per million input/output tokens.Epoch AI; OpenAIMEDIUM
2025-10-06ChatGPT at 800M weekly usersConsumer AI reaches roughly a tenth of humanity weekly. 800M weekly active users.OpenAI DevDay via TechCrunchMEDIUM
2025-11-24Claude Opus 4.5 passes 80% on SWE-bench VerifiedAutonomous software-engineering benchmark crosses 80%. 80.9% (vendor-reported).AnthropicMEDIUM
2026-02ARC-AGI-2 above 80%The benchmark launched in Mar 2025 to replace near-solved ARC-AGI-1 is largely solved within a year. 84.6% (Gemini 3 Deep Think upgrade, ARC Prize-verified, ~$13.62 per task).ARC Prize via OfficeChaiMEDIUM
2026-03-31OpenAI raises $122B at $852BLargest private funding round on record. $122B at $852B post-money.OpenAI; BloombergMEDIUM
2026-04-07Claude Mythos Preview withheld from general release; Project GlasswingAnthropic announced its most capable model but restricted it to defensive-security partners because of its cyber capabilities: the first frontier model deliberately withheld from general availability for capability reasons. 12 launch organisations (incl. Anthropic) plus 40+ more; up to $100M in usage credits and $4M in donations to open-source security groups; $25 / $125 per M tokens after credits.Anthropic, Project Glasswing (opened 12 Nov 2026)HIGH
2026-04-13AI Index 2026: corporate AI investment more than doublesStanford HAI measures the 2025 investment surge. Global corporate AI investment $581.7B in 2025 (+130%); private investment $344.7B.Stanford HAI AI Index 2026MEDIUM
2026-06-09Claude Fable 5 / Mythos 5 launched; suspended 12 Jun under US export controls, restored 1 JulAnthropic released Fable 5 generally and Mythos 5 to Glasswing partners on 9 Jun 2026, suspended both for all customers on 12 Jun after a US export-control directive, and redeployed both on 1 Jul (Mythos 5 for a set of US organisations only). $10 / $50 per million input/output tokens; access suspended for 19 days (12 Jun–1 Jul).Anthropic launch post with 12 Jun and 1 Jul updates (opened 12 Nov 2026)HIGH
2026-07Perfect scores at IMO 2026Several AI systems reported 42/42 at the Shanghai olympiad, saturating the contest as a benchmark. 42/42 reported for multiple systems (trade press).Trade press via Digital AppliedMEDIUM
2026-08-31ChatGPT passes 1 billion weekly usersFirst AI product above one billion weekly users. 1B+ weekly active users (OpenAI).OpenAIMEDIUM
2026-09-01SWE-bench Verified saturatedIndependent leaderboard stops running the benchmark on new releases. 97.0% (Claude Opus 5, Vals.ai harness).Vals.aiMEDIUM
2026-09-01Claude Fable 5.1 / Mythos 5.1Same model with two safeguard levels: Fable 5.1 generally available, Mythos 5.1 only through trusted-access programmes. Fable 5.1 lists at $10 / $50 per M tokens.Anthropic, Fable 5.1 / Mythos 5.1 page (opened 12 Nov 2026; page dated SeptemberHIGH
2026-09-03GPT-6 Astra: largest estimated training runEpoch’s largest training-compute estimate to date. ~9.4×10^26 FLOP (Epoch, ‘Likely’; CI ~5e26–2e27), ~45× GPT-4; ≥100,000 GB200s, ~233 MW.Epoch AI Notable AI Models CSVHIGH
2026-10-26Anthropic files S-1First frontier lab to file publicly for an IPO. H1 2026 revenue $16.3B; Q2 2026 first operating profit ($974M).Anthropic S-1 as reportedMEDIUM
2026-11-06OpenAI valued at $1.25 trillionFirst private AI lab valued above $1 trillion. First close of ~$11B at a reported $1.25T.Bloomberg, Reuters as reportedMEDIUM
Fact check

Numbers to stop quoting

Widely repeatedWhat the evidence supports
“The term ‘artificial intelligence’ was coined at the 1956 Dartmouth conference.”It first appears in the Dartmouth proposal dated 31 August 1955 (McCarthy, Minsky, Rochester, Shannon). The workshop itself ran in summer 1956 — about 18 June to 17 August per Solomonoff’s notes, six weeks in other accounts — and attendance varied; the proposal asked for a ‘2 month, 10 man study’.
“The AI winters were 1974–1980 and 1987–1993.”These are conventions. UK cuts followed the Lighthill report in 1973; DARPA’s cuts came in 1973–74. For the second winter Wikipedia gives 1987–2000 while other histories use 1987–1993. Say ‘roughly’ and name the defining events (DARPA cuts; Lisp-machine collapse).
“Hinton invented backpropagation in 1986.”Rumelhart, Hinton and Williams popularised it for neural networks (Nature, 9 Oct 1986). The method appeared earlier: Linnainmaa’s 1970 master’s thesis (reverse-mode differentiation), Werbos’s 1974 Harvard thesis, Parker (1985).
“Minsky and Papert’s ‘Perceptrons’ (1969) killed neural-network research and caused the AI winter.”It was one factor among several. The 1969 Mansfield Amendment redirected DARPA money to mission-oriented work, and the Lighthill report (1973) hit UK funding. Sources do not establish a single cause.
“Deep Blue beat Kasparov in 1996.”Kasparov won the February 1996 match 4–2 (Deep Blue won one game, on 10 Feb 1996). Deep Blue won the May 1997 rematch 3½–2½, with the decisive game on 11 May 1997.
“A chatbot passed the Turing test in 2014 (Eugene Goostman).”At a University of Reading event on 7 June 2014, 33% of 30 judges were fooled in five-minute chats by a bot posing as a 13-year-old Ukrainian boy. That clears an organiser-chosen 30% bar, not a standard Turing set: his 1950 paper made a prediction (no more than a 70% chance of correct identification after five minutes by about 2000), not a pass criterion.
“Humans make 5.1% errors on ImageNet, so AI beat humans in 2015.”5.1% is one trained annotator (Andrej Karpathy) on a sample of validation images in 2014; two-thirds of his errors came from not knowing the 1,000 classes. ResNet’s 3.57% (Dec 2015) beat that single estimate, not ‘humans’.
“ChatGPT reached 100 million users in two months — the fastest-growing app ever.”The 100M figure was a UBS estimate of monthly users in January 2023 from Similarweb web-traffic data, not an OpenAI disclosure; OpenAI’s own first 100M figure was weekly users in November 2023. Threads passed 100 million sign-ups in five days in July 2023 — sign-ups, not active users, so the two records are not comparable.
“GPT-4 has 1.76 trillion parameters.”OpenAI has never disclosed it. 1.76T traces to George Hotz’s speculation; Semafor reported ~1T and SemiAnalysis ~1.8T. The mixture-of-experts design is also unconfirmed. Stanford’s AI Index 2026 notes frontier labs have stopped reporting parameter counts. Note that even Epoch’s database stores 1.8T in its Parameters field, explicitly flagged in its notes as a rumour — do not cite Epoch as confirmation.
“GPT-4 scored in the top 10% on the bar exam.”OpenAI’s percentile compared GPT-4 with February Illinois bar takers, a group skewed towards repeat takers. A 2024 re-evaluation (Artificial Intelligence and Law) found a much lower percentile against first-time or all July takers — Live Science summarised it as not breaking the 70th percentile.
“AI training compute doubles every 3.4 months.”That was OpenAI’s 2018 estimate for 2012–2018. Epoch’s longer series gives ~5.7–6 months for the deep-learning era, about 4–5× a year for notable models since 2010 and ~5× a year for frontier language models since 2020. Before 2010 the doubling time was about 20 months.
“OpenAI’s o3 solved 25% of FrontierMath.”25.2% was OpenAI’s own December 2024 run, probably with more compute or scaffolding. Epoch’s independent test of the released o3 in April 2025 found about 10%. OpenAI had funded the benchmark and had access to many problems, which Epoch disclosed only later.
“AlphaGo Zero used 3.41×10^23 FLOP and proves compute grows 10× a year.”3.41e23 is OpenAI’s 2018 estimate and counts the compute spent generating self-play games. Epoch’s database now counts only direct network training: 6.5×10^20 FLOP for AlphaGo Zero (‘Confident’) and 3.4×10^20 for AlphaGo Master. Quote the basis with the number; neither figure supports a ’10× a year’ trend claim.
“Google paid $500 million for DeepMind.”Terms were not disclosed. Reports ranged from $400 million (Re/code) to ‘more than $500 million’ (The Information).
“XCON saved DEC $40 million a year.”Most accounts put R1/XCON’s savings at about $25 million a year; $40 million was DEC’s figure for XCON plus related expert systems.
“Training frontier models is getting cheaper (AI Index 2026: −30%, ~$120M).”That figure comes from one secondary summary and could not be checked in the report (hai.stanford.edu unreachable). Epoch’s trends page says frontier language-model training cost has risen ~3.5× a year since 2020, and its data put 2025 runs at $218M–$388M amortised (2023 USD) and Grok 4 at ~$490M nominal. Quote Epoch until the AI Index text is verified.
“The US produced 61 notable AI models in 2023.”61 is the AI Index 2024 count; the AI Index 2025 restated 2023 as 70. Quote the edition. (2024: 40; 2025: 50 or 59 depending on the 2026 PDF vs web page.)
“AI inference costs have fallen 280-fold.”Stanford’s figure covers GPT-3.5-level capability from $20 (Nov 2022) to $0.07 (Oct 2024) per million tokens; it is not the price of frontier AI. · detail
“DeepSeek trained a frontier model for $5.6 million.”$5.58M was DeepSeek’s stated GPU-rental cost for V3’s final pre-training run only — excluding research, earlier runs, staff and hardware — and is not comparable with Epoch’s ~$490M estimate for Grok 4.
“Claude Fable 5 was withdrawn within a week of launch (and is unavailable).”Anthropic suspended Fable 5 and Mythos 5 on 12 June 2026, three days after the 9 June launch, under a US export-control directive, and restored both on 1 July 2026 (Mythos 5 for a set of US organisations only). Fable 5.1 has been generally available since September 2026. · detail
Definitions

Terms used in this index

Training compute (FLOP)
Total floating-point operations used to train a model; Epoch AI estimates it from hardware, training time and utilisation, or from parameters and training data.
Notable model
A model in Epoch AI’s database that meets its notability criteria (state of the art, highly cited, historically significant or widely used).
Amortised training cost
Hardware depreciation plus energy for the final training run, in constant dollars (Epoch AI method).
Cloud-rental training cost
The price of renting the same compute from a cloud provider; typically two to four times the amortised cost.
Constant 2021 dollars
Values adjusted for inflation to 2021 prices, so years can be compared directly.
Weekly active users
People who used a service at least once in a seven-day period; not comparable with monthly active users.

Frequently asked

What is the AI Development Index?

Report AI’s AI Development Index is a dated, sourced record of how AI has developed since 1950, tracked through six measures: training compute, capability (benchmarks), cost, capital, adoption, and risk and resources. It is updated quarterly and the full dataset is free to download as a CSV.

Why not combine the measures into one score?

A single score would need weights that are arbitrary and would hide the fact that the measures move at very different speeds. Keeping them separate lets readers see, for example, that prices for a fixed capability fall while frontier training costs rise.

What is the largest AI training run so far?

Epoch AI’s estimate for OpenAI’s GPT-6 Astra, released in September 2026, is about 9.4×10²⁶ floating-point operations, roughly 45 times GPT-4. Epoch rates the estimate ‘Likely’, with a range of about 5×10²⁶ to 2×10²⁷.

How fast is AI compute growing?

Epoch AI estimates that training compute for frontier language models has grown about 5× a year since 2020 (90% interval 4–6×), and that compute for notable models doubled about every six months after 2010, against about every 20 months before.

How much is invested in AI each year?

The Stanford HAI AI Index 2026 puts global corporate AI investment at $581.7 billion in 2025 in current dollars, of which private investment was $344.7 billion. In constant 2021 dollars, Our World in Data’s series shows corporate investment of about $490 billion in 2025, up from about $17 billion in 2013.

Methodology & sources

Figures are compiled from company disclosures, regulators, research firms and reputable reporting, and cross-checked against Report AI’s own index pages. Each value carries its as-of date, source and a confidence rating: HIGH where the primary source was read, MEDIUM where it rests on consistent reputable secondary reporting (most figures, because many primary pages cannot be opened from our research environment), DERIVED where it is our own arithmetic, and CONTESTED where credible sources disagree. A year means the calendar year the figure describes; 2026 values are the latest available and marked “to date”. Weekly and monthly user counts, and forecasts and outcomes, are never compared as if they were the same measure. The AI Development Index draws on Epoch AI’s Notable AI Models dataset (downloaded 12 November 2026) for compute, parameters and training costs; Our World in Data’s republication of the Stanford HAI AI Index for investment and model counts; the AI Index for prices, benchmarks, incidents, publications and patents; McKinsey’s Global Survey on AI for organisational adoption; OpenAI for ChatGPT users; and the IEA, LBNL and Gartner for electricity. The decade scoreboard, growth multiples and the 187-fold price fall are Report AI calculations (DERIVED). Values that could not be checked against a primary source in this edition are rated MEDIUM; the AI Index 2026 summary claim of about $120 million for the largest mid-2025 training run is CONTESTED because Epoch’s estimates for 2025 runs are higher. These reports compile and verify secondary data with original analysis; they are not original surveys. See our editorial standards, corrections log and fact-check register.