Level : Intermediate
AI in Numbers: The Essential Charts
These 8 interactive charts sum up how AI is evolving, based on public data: performance, prices, models, adoption, investment and compute. Each one cites its source and date, and the guide below gives you the keys to read them without jumping to conclusions.
Model progress on benchmarks
Best raw score achieved by an AI model on each benchmark, by the model's release quarter (running record). Click the legend to hide or show a benchmark.
On GPQA Diamond, the best score rose from 35.7% (Q1 2023) to 96.4% (Q3 2026).
Show the data
| Quarter | GPQA Diamond | OTIS Mock AIME | FrontierMath | SWE-bench Verified | FrontierCode | Terminal-Bench | APEX-Agents |
|---|---|---|---|---|---|---|---|
| Q1 2023 | 35.7 % | 0.6 % | — | — | — | — | — |
| Q2 2023 | 35.7 % | 1.1 % | — | — | — | — | — |
| Q3 2023 | 35.7 % | 2.5 % | — | — | — | — | — |
| Q4 2023 | 42.4 % | 2.5 % | — | — | — | — | — |
| Q1 2024 | 47.2 % | 4.7 % | 0 % | — | — | — | — |
| Q2 2024 | 54 % | 6.8 % | 0.7 % | — | — | — | — |
| Q3 2024 | 62.4 % | 46.9 % | 0.7 % | — | — | — | — |
| Q4 2024 | 76.8 % | 73.3 % | 14.7 % | 31 % | — | — | — |
| Q1 2025 | 83.8 % | 76.9 % | 18.6 % | 61 % | — | — | — |
| Q2 2025 | 85.3 % | 84.7 % | 36.1 % | 70.7 % | — | 32.6 % | — |
| Q3 2025 | 87 % | 91.4 % | 55.4 % | 73.6 % | — | 49.6 % | 4.4 % |
| Q4 2025 | 92.6 % | 96.1 % | 74 % | 76.7 % | — | 69.4 % | 21.3 % |
| Q1 2026 | 94.6 % | 97.8 % | 82.5 % | 78.7 % | 27 % | 81.8 % | 52.4 % |
| Q2 2026 | 94.6 % | 100 % | 87.7 % | 83.5 % | 53.5 % | 84.7 % | 63.6 % |
| Q3 2026 | 96.4 % | 100 % | 93.7 % | 83.5 % | 54.6 % | 84.7 % | 82.4 % |
| Q4 2026 | 96.4 % | 100 % | 93.7 % | 83.5 % | 54.6 % | 84.7 % | 82.4 % |
The price of a given level of performance
Lowest public price (US dollars per million tokens, 3 input tokens for 1 output token) among models already released that reach the stated GPQA Diamond threshold. Logarithmic scale. Current prices from models.dev (developer's price, otherwise the hosts' median), not historical prices: only models still offered and evaluated by Epoch AI count (145 models).
For at least 50% on GPQA Diamond, the lowest price falls from $0.49 (Q4 2024, DeepSeek-V3) to $0.055 (Q3 2026, Qwen3.7 Flash), about 9 times cheaper.
Show the data
| Quarter | GPQA Diamond ≥ 50% | $ / M | GPQA Diamond ≥ 70% | $ / M |
|---|---|---|---|---|
| Q4 2024 | DeepSeek-V3 (56.5 %) | $0.49 | o1 (76.8 %) | $26.25 |
| Q1 2025 | QwQ-32B (65.3 %) | $0.40 | DeepSeek-R1 (71.7 %) | $1.22 |
| Q2 2025 | Qwen3-8B (56.8 %) | $0.098 | Qwen3-235B-A22B (70.7 %) | $0.35 |
| Q3 2025 | Qwen3-8B (56.8 %) | $0.098 | gpt-oss-120b (75.8 %) | $0.26 |
| Q4 2025 | Qwen3-8B (56.8 %) | $0.098 | gpt-oss-120b (75.8 %) | $0.26 |
| Q1 2026 | Qwen3-8B (56.8 %) | $0.098 | Qwen3.5-9B (79 %) | $0.11 |
| Q2 2026 | Qwen3-8B (56.8 %) | $0.098 | Qwen3.5-9B (79 %) | $0.11 |
| Q3 2026 | Qwen3.7 Flash (82.3 %) | $0.055 | Qwen3.7 Flash (82.3 %) | $0.055 |
| Q4 2026 | Qwen3.7 Flash (82.3 %) | $0.055 | Qwen3.7 Flash (82.3 %) | $0.055 |
Notable AI models by year and country
Number of models Epoch AI considers notable, by publication year and the country of the first organisation listed. Click the legend to hide a country.
In 2025, Epoch AI lists 107 notable models, including 62 in the United States and 34 in China.
Show the data
| Year | United States | China | United Kingdom | South Korea | Canada | Other countries | Total |
|---|---|---|---|---|---|---|---|
| 2012 | 5 | 0 | 0 | 0 | 6 | 2 | 13 |
| 2013 | 12 | 0 | 2 | 0 | 1 | 4 | 19 |
| 2014 | 13 | 2 | 7 | 0 | 4 | 7 | 33 |
| 2015 | 18 | 3 | 4 | 0 | 0 | 4 | 29 |
| 2016 | 32 | 4 | 5 | 0 | 1 | 7 | 49 |
| 2017 | 23 | 9 | 10 | 3 | 3 | 7 | 55 |
| 2018 | 26 | 1 | 4 | 1 | 0 | 5 | 37 |
| 2019 | 51 | 5 | 7 | 1 | 0 | 4 | 68 |
| 2020 | 34 | 6 | 4 | 1 | 1 | 4 | 50 |
| 2021 | 49 | 18 | 4 | 1 | 1 | 6 | 79 |
| 2022 | 61 | 10 | 13 | 3 | 1 | 3 | 91 |
| 2023 | 81 | 23 | 2 | 1 | 1 | 11 | 119 |
| 2024 | 55 | 30 | 1 | 1 | 1 | 10 | 98 |
| 2025 | 62 | 34 | 1 | 8 | 0 | 2 | 107 |
| 2026* | 49 | 29 | 0 | 5 | 0 | 1 | 84 |
European businesses using AI
Share of enterprises with 10 or more employees using at least one AI technology, by country. The European Union average is highlighted.
In 2021, 7.7% of EU enterprises use AI, from 0.9% (Serbia) to 23.9% (Denmark) depending on the country.
In 2023, 8.1% of EU enterprises use AI, from 1.5% (Romania) to 15.2% (Denmark) depending on the country.
In 2024, 13.5% of EU enterprises use AI, from 3.1% (Romania) to 27.6% (Denmark) depending on the country.
In 2025, 20% of EU enterprises use AI, from 5.2% (Romania) to 42% (Denmark) depending on the country.
Show the data
| Country | 2021 | 2023 | 2024 | 2025 |
|---|---|---|---|---|
| EU average (27) | 7.7 % | 8.1 % | 13.5 % | 20 % |
| Belgium | 10.3 % | 13.8 % | 24.7 % | 34.5 % |
| Bulgaria | 3.3 % | 3.6 % | 6.5 % | 8.6 % |
| Czechia | 4.5 % | 5.9 % | 11.3 % | 17.6 % |
| Denmark | 23.9 % | 15.2 % | 27.6 % | 42 % |
| Germany | 10.6 % | 11.6 % | 19.8 % | 26 % |
| Estonia | 2.8 % | 5.2 % | 13.9 % | 23.4 % |
| Ireland | 7.9 % | 8 % | 14.9 % | 19.6 % |
| Greece | 2.6 % | 4 % | 9.8 % | 8.9 % |
| Spain | 7.7 % | 9.2 % | 11.3 % | 20.3 % |
| France | 6.7 % | 5.9 % | 9.9 % | 18.2 % |
| Croatia | 8.7 % | 7.9 % | 11.8 % | 15.2 % |
| Italy | 6.2 % | 5.1 % | 8.2 % | 16.4 % |
| Cyprus | 2.6 % | 4.7 % | 7.9 % | 9.3 % |
| Latvia | 3.7 % | 4.5 % | 8.8 % | 12.2 % |
| Lithuania | 4.5 % | 4.9 % | 8.8 % | 21.3 % |
| Luxembourg | 13 % | 14.5 % | 23.7 % | 33.6 % |
| Hungary | 3 % | 3.7 % | 7.4 % | 10.4 % |
| Malta | 10.2 % | 13.2 % | 17.3 % | 21.5 % |
| Netherlands | 13.1 % | 14.1 % | 23.1 % | 33.2 % |
| Austria | 8.8 % | 10.8 % | 20.3 % | 30 % |
| Poland | 2.9 % | 3.7 % | 5.9 % | 8.4 % |
| Portugal | 7.2 % | 7.9 % | 8.6 % | 11.5 % |
| Romania | 1.4 % | 1.5 % | 3.1 % | 5.2 % |
| Slovenia | 11.7 % | 11.4 % | 20.9 % | 21.6 % |
| Slovakia | 5.2 % | 7 % | 10.8 % | 18 % |
| Finland | 15.8 % | 15.1 % | 24.4 % | 37.8 % |
| Sweden | 9.9 % | 10.4 % | 25.1 % | 35 % |
| Norway | 10.8 % | 9.2 % | 20.8 % | 28.9 % |
| Bosnia & Herzegovina | 2.1 % | 5.3 % | 6.4 % | 10.8 % |
| Montenegro | 3.3 % | 5.6 % | 7.9 % | 10.1 % |
| Albania | 3.7 % | — | 8.9 % | 9 % |
| Serbia | 0.9 % | 1.8 % | 7 % | 10.1 % |
| Türkiye | 2.7 % | 5.5 % | 4.4 % | 7.4 % |
Private investment in AI
Annual private investment in AI companies, in US dollars, according to Our World in Data.
In 2025, private investment in AI reached $290bn worldwide, including $241bn in the United States.
Show the data
| Year | United States | China | Europe | World |
|---|---|---|---|---|
| 2013 | $4.2bn | $0.7bn | $0.8bn | $6bn |
| 2014 | $8bn | $0.8bn | $1.5bn | $11bn |
| 2015 | $11bn | $2.4bn | $1bn | $15bn |
| 2016 | $11bn | $5.1bn | $1.4bn | $19bn |
| 2017 | $16bn | $7.3bn | $2.9bn | $28bn |
| 2018 | $23bn | $15bn | $3.2bn | $47bn |
| 2019 | $34bn | $15bn | $6.7bn | $62bn |
| 2020 | $46bn | $18bn | $6.5bn | $77bn |
| 2021 | $87bn | $26bn | $15bn | $145bn |
| 2022 | $57bn | $14bn | $14bn | $105bn |
| 2023 | $64bn | $8.4bn | $11bn | $93bn |
| 2024 | $95bn | $8.1bn | $17bn | $131bn |
| 2025 | $241bn | $10bn | $18bn | $290bn |
Global investment in generative AI
Annual private investment in generative AI companies worldwide.
From $0.9bn in 2019 to $144bn in 2025.
Show the data
| Year | Value |
|---|---|
| 2019 | $0.9bn |
| 2020 | $2.5bn |
| 2021 | $5.3bn |
| 2022 | $3.7bn |
| 2023 | $25bn |
| 2024 | $44bn |
| 2025 | $144bn |
Compute used to train notable AI systems
Each dot is an AI system, placed by publication date and training compute (in petaFLOP, i.e. 10¹⁵ operations). Logarithmic scale: each gridline multiplies by 10,000. Darker, keyboard-focusable dots mark a new record.
536 systems, from 4 × 10⁻¹⁴ petaFLOP (Theseus, 1950) to 5 × 10¹¹ petaFLOP (Grok 4, 2025).
Show the data : Successive records
| Model | Year | petaFLOP |
|---|---|---|
| Theseus | July 2, 1950 | 4 × 10⁻¹⁴ |
| Perceptron Mark I | January 1, 1957 | 6.9 × 10⁻¹⁰ |
| Pandemonium (morse) | February 1, 1959 | 6 × 10⁻⁷ |
| Perceptron (1960) | March 30, 1960 | 7.2 × 10⁻⁷ |
| NetTalk (dictionary) | June 6, 1987 | 2.8 × 10⁻⁵ |
| NetTalk (transcription) | June 6, 1987 | 2.8 × 10⁻⁵ |
| Handwritten digit recognition network | November 27, 1989 | 1.8 × 10⁻⁴ |
| Zip CNN | December 1, 1989 | 1.5 × 10⁻³ |
| TD-Gammon | May 1, 1992 | 0.02 |
| Predictive Coding NN | December 2, 1994 | 0.02 |
| LSTM | November 15, 1997 | 0.03 |
| Neural LM | November 28, 2000 | 6.34 |
| KN-LM | June 22, 2007 | 773.04 |
| SB-LM | June 22, 2007 | 1,449.45 |
| DistBelief NNLM | January 16, 2013 | 2,612.74 |
| SPPNet | June 18, 2014 | 3,411.07 |
| VGG16 | September 4, 2014 | 1.2 × 10⁴ |
| Seq2Seq LSTM | September 10, 2014 | 5.6 × 10⁴ |
| SNM-skip | December 3, 2014 | 3 × 10⁵ |
| AlphaGo Fan | October 1, 2015 | 3.8 × 10⁵ |
| AlphaGo Lee | January 27, 2016 | 1.9 × 10⁶ |
| GNMT | September 26, 2016 | 6.6 × 10⁶ |
| ResNeXt-101 32x48d | May 2, 2018 | 8.7 × 10⁶ |
| Megatron-BERT | September 17, 2019 | 2.2 × 10⁷ |
| T5-11B | October 23, 2019 | 3.3 × 10⁷ |
| AlphaStar | October 30, 2019 | 1.1 × 10⁸ |
| Meena | January 28, 2020 | 1.1 × 10⁸ |
| GPT-3 175B (davinci) | May 28, 2020 | 3.1 × 10⁸ |
| Jurassic-1-Jumbo | August 11, 2021 | 3.7 × 10⁸ |
| FLAN 137B | September 3, 2021 | 2 × 10⁹ |
| GPT-3.5 (davinci-002) | March 15, 2022 | 2.6 × 10⁹ |
| Minerva (540B) | June 29, 2022 | 2.7 × 10⁹ |
| GPT-4 (Mar 2023) | March 15, 2023 | 2.1 × 10¹⁰ |
| Gemini 1.0 Ultra | December 6, 2023 | 5 × 10¹⁰ |
| Grok 3 | February 17, 2025 | 3.5 × 10¹¹ |
| GPT-4.5 | February 27, 2025 | 3.8 × 10¹¹ |
| Grok 4 | July 9, 2025 | 5 × 10¹¹ |
Europeans using generative AI
Share of people aged 16 to 74 who used a generative AI tool, by country, in 2025.
In 2025, the share ranges from 17.2% (Türkiye) to 56.3% (Norway).
Show the data
| Country | % |
|---|---|
| Norway | 56.3 % |
| Denmark | 48.4 % |
| Switzerland | 47 % |
| Estonia | 46.6 % |
| Malta | 46.5 % |
| Finland | 46.3 % |
| Ireland | 44.9 % |
| Netherlands | 44.7 % |
| Cyprus | 44.2 % |
| Greece | 44.1 % |
| Luxembourg | 42.5 % |
| Belgium | 42 % |
| Sweden | 42 % |
| Austria | 39.4 % |
| Portugal | 38.7 % |
| Spain | 37.9 % |
| Slovenia | 37.6 % |
| France | 37.5 % |
| Lithuania | 36.9 % |
| Czechia | 35.4 % |
| Latvia | 33.4 % |
| Germany | 32.3 % |
| Slovakia | 30.8 % |
| Hungary | 29.6 % |
| Croatia | 27.5 % |
| Poland | 22.7 % |
| Bulgaria | 22.5 % |
| North Macedonia | 22 % |
| Bosnia & Herzegovina | 20.3 % |
| Italy | 19.9 % |
| Serbia | 18.6 % |
| Romania | 17.8 % |
| Türkiye | 17.2 % |
What the charts show
The charts above cover the main dimensions of AI:
- benchmark progress, the standardized test suites used to compare models;
- falling prices for reaching a given level of performance;
- the number of notable models released each year and how they split by country;
- AI adoption by European businesses, country by country, based on Eurostat;
- investment in the sector;
- training compute, meaning the amount of computation needed to build a model;
- how many Europeans use generative AI, according to Eurostat.
Hover over or tap a curve to see exact values. This text does not repeat them: they change every time the sources are updated.
Reading a logarithmic scale
Several charts use a logarithmic scale on the vertical axis. On a regular scale, each gridline adds the same amount. On a logarithmic scale, each gridline multiplies the value by the same factor.
Why? Because some quantities, such as training compute, have changed by enormous proportions. On a regular scale, older models would be squashed at the bottom and invisible. On a log scale, you can take in the whole history at a glance.
Two reading habits:
- a straight rising line means growth at a steady rate, where the value doubles at regular intervals, not a modest increase;
- a small vertical gap can hide a huge real difference. Always read the gridlines before comparing two points.
Correlation is not causation
Two curves rising together do not prove that one drives the other. For instance, if AI adoption is higher in countries where businesses are already highly digitized, that does not mean AI caused the digitization, or the reverse. A third factor, such as the structure of the economy or company size, may explain both. The charts show associations; causes require dedicated studies.
What the data says, and what it does not
- Benchmarks: a rising score shows a model does better on a specific test, not that it is better at everything. Some tests eventually become “saturated,” and test questions may have leaked into training data.
- Prices: the drop applies to a given level of performance. Cutting-edge models can stay expensive, and your real bill depends on how you use them.
- Notable models: the count relies on Epoch AI’s selection, which reflects what has been published and documented. The country assigned is that of the organizations, not necessarily where the work was done.
- Adoption in Europe: Eurostat’s data is based on self-reported business surveys. What a company counts as “using AI” can vary.
- Investment and compute: these are often estimates built from partial information. They point to an order of magnitude and a trend more than an exact value.
Sources and method
The data comes from Epoch AI, Our World in Data and Eurostat, under their respective licenses. Our methodology explains how we use and update it.
To put these trends in historical context, browse the AI timeline. To compare today’s models in practice, see the AI ranking or the compare models tool. Stuck on a term? The glossary has you covered.
Frequently asked questions
Where does the data in these charts come from?
It comes from well-known public sources: Epoch AI, a research institute that tracks the development of AI; Our World in Data, which compiles global statistics; and Eurostat, the statistical office of the European Union. The source and date appear under each chart.
Why do some axes not go up in even steps?
They use a logarithmic scale: each gridline multiplies the value by the same factor instead of adding the same amount. This lets one chart show both very small and very large quantities, such as the compute needed to train models from different eras.
How up to date are the charts?
The site is rebuilt every week and pulls in the latest versions published by each source. Each organization updates its data on its own schedule, though, so rely on the date shown under each chart.
Can these numbers predict the future of AI?
No. They describe past trends, with their uncertainties. Extending a curve is tempting, but nothing guarantees the pace will hold: technical, economic or regulatory limits can change it.