AI progress
These charts track measurable capabilities on specific tests and the computing used to train models. They are not a measure of general intelligence, and they do not forecast what comes next.
Data retrieved . Each chart gives the date its data runs through.
Training compute of notable models
Training compute is the number of calculations (floating-point operations, or FLOP) used to train a model. Each point is one model from Epoch AI's database of notable models: those that improved on a recognized benchmark, are highly cited, are historically significant, or are widely used. It is a curated list, not every model.
The vertical axis is a log scale: each gridline is 10 times the one below. A straight line on this scale means growth by the same multiple each year, not by the same amount.
Training compute of notable AI models, log scale
540 of the 1,072 models in Epoch's file have a training compute figure; only those are shown. Developers often do not report compute, so many figures are Epoch's own estimates: the fill of each point shows Epoch's confidence in it. Rings mark models tied to milestones below.
Left and Right arrows move from point to point in date order; Page Up and Page Down jump a tenth of the way later or earlier; Home and End go to the first and last point; Escape clears.
Left and Right arrows move from point to point in date order; Page Up and Page Down jump a tenth of the way later or earlier; Home and End go to the first and last point; Escape clears.
Left and Right arrows move from point to point in date order; Page Up and Page Down jump a tenth of the way later or earlier; Home and End go to the first and last point; Escape clears.
Data table
| Year | Model | Organization | Domain | Training compute (FLOP) | Epoch's confidence |
|---|---|---|---|---|---|
| 1950 | Theseus | Bell Laboratories | Robotics and other | 40 | Confident |
| 1957 | Perceptron Mark I | Cornell Aeronautical Laboratory, Cornell University | Robotics and other | 6.9 × 10⁵ | Likely |
| 1959 | Pandemonium (morse) | Massachusetts Institute of Technology (MIT) | Language | 6.0 × 10⁸ | Speculative |
| 1959 | Samuel Neural Checkers | IBM | Games | 4.3 × 10⁸ | Likely |
| 1960 | Perceptron (1960) | Cornell Aeronautical Laboratory | Vision, images and video | 7.2 × 10⁸ | Speculative |
| 1960 | ADALINE | Stanford University | Vision, images and video | 6.6 × 10³ | Confident |
| 1962 | Linear Decision Functions | Bell Laboratories | Biology and other sciences | 1.6 × 10⁶ | Speculative |
| 1963 | Print Recognition Logic | IBM | Vision, images and video | 2.3 × 10⁷ | Speculative |
| 1965 | Heuristic Reinforcement Learning | Purdue University | Robotics and other | 1.1 × 10⁶ | Speculative |
| 1966 | LTE speaker verification system | IBM | Speech and audio | 1.1 × 10⁸ | Likely |
| 1975 | Cognitron | Biological Cybernetics | Robotics and other | 5.2 × 10⁶ | Confident |
| 1980 | Neocognitron | NHK Broadcasting Science Research Laboratories | Vision, images and video | 2.7 × 10⁸ | Confident |
| 1983 | ASE+ACE | University of Massachusetts Amherst | Robotics and other | 3.2 × 10⁸ | Likely |
| 1986 | MLP with back-propagation | University of California San Diego, Carnegie Mellon University (CMU) | Biology and other sciences | 6.7 × 10⁸ | Confident |
| 1986 | Distributed representation NN | Carnegie Mellon University (CMU) | Robotics and other | 3.9 × 10⁸ | Confident |
| 1987 | NetTalk (transcription) | Princeton University | Speech and audio | 2.8 × 10¹⁰ | Confident |
| 1987 | NetTalk (dictionary) | Princeton University | Speech and audio | 2.8 × 10¹⁰ | Confident |
| 1987 | Translation-invariant MLP | Carnegie Mellon University (CMU) | Robotics and other | 1.8 × 10¹⁰ | Confident |
| 1988 | MLN-ASR | McGill University | Speech and audio | 3.0 × 10⁸ | Confident |
| 1989 | Zip CNN | AT&T, Bell Laboratories | Vision, images and video | 1.5 × 10¹² | Confident |
| 1989 | Handwritten digit recognition network | AT&T | Vision, images and video | 1.8 × 10¹¹ | Confident |
| 1989 | Invariant image recognition | Complutense University of Madrid | Vision, images and video | 2.7 × 10¹⁰ | Confident |
| 1990 | SexNet compression | Vision, images and video | 7.9 × 10¹⁰ | Confident | |
| 1990 | NETtalk reimplementation | Oregon State University | Speech and audio | 3.6 × 10¹⁰ | Confident |
| 1990 | Bankruptcy-NN | Robotics and other | 3.1 × 10⁹ | Confident | |
| 1991 | Weight Decay | Speech and audio | 7.5 × 10¹⁰ | Confident | |
| 1992 | TD-Gammon | IBM | Games | 1.8 × 10¹³ | Speculative |
| 1992 | Cancer drug mechanism prediction | National Cancer Institute | Biology and other sciences | 5.3 × 10⁷ | Likely |
| 1993 | Siamese-TDNN | Bell Laboratories | Vision, images and video | 1.3 × 10¹³ | Likely |
| 1993 | ANN Eye Tracker | Vision, images and video | 1.8 × 10¹⁰ | Confident | |
| 1994 | Predictive Coding NN | Technical University of Munich | Language | 1.9 × 10¹³ | Confident |
| 1994 | NeuroChess | Games | 8.6 × 10¹¹ | Speculative | |
| 1994 | Mixture of linear models | Vision, images and video | 4.5 × 10¹¹ | Likely | |
| 1995 | LISSOM | University of Texas at Austin | Vision, images and video | 2.0 × 10¹¹ | Likely |
| 1996 | MUSIC perceptron | Vision, images and video | 8.8 × 10¹¹ | Confident | |
| 1996 | System 11 | Carnegie Mellon University (CMU) | Vision, images and video | 2.6 × 10¹⁰ | Confident |
| 1997 | LSTM | Technical University of Munich | Language | 3.2 × 10¹³ | Confident |
| 1997 | SOM-CNN | Vision, images and video | 3.1 × 10¹⁰ | Confident | |
| 1998 | LeNet-5 | AT&T | Vision, images and video | 2.8 × 10¹² | Confident |
| 1999 | RECONTRA-categorized | Language | 8.0 × 10¹² | Likely | |
| 1999 | RECONTRA-uncategorized | Language | 3.9 × 10¹² | Likely | |
| 2000 | Neural LM | University of Montreal / Université de Montréal | Language | 6.3 × 10¹⁵ | Confident |
| 2000 | PoE MNIST | University College London (UCL) | Vision, images and video | 5.2 × 10¹³ | Confident |
| 2001 | Decision tree (classification) | Mitsubishi Electric Research Labs, Compaq CRL | Vision, images and video | 6.3 × 10¹³ | Likely |
| 2003 | NPLM (AP News) | University of Montreal / Université de Montréal | Language | 1.7 × 10¹⁵ | Confident |
| 2003 | NPLM (Brown) | University of Montreal / Université de Montréal | Language | 1.3 × 10¹⁴ | Confident |
| 2004 | LMICA | Vision, images and video | 2.8 × 10¹⁵ | Confident | |
| 2004 | Invariant CNN | New York University (NYU) | Vision, images and video | 9.7 × 10¹¹ | Confident |
| 2005 | Hierarchical LM | Language | 1.2 × 10¹⁴ | Confident | |
| 2005 | RankNet | Microsoft Research, Microsoft | Robotics and other | 3.5 × 10¹² | Confident |
| 2006 | SVM-CNN | New York University (NYU) | Vision, images and video | 7.5 × 10¹⁴ | Confident |
| 2007 | SB-LM | Language | 1.4 × 10¹⁸ | Likely | |
| 2007 | KN-LM | Language | 7.7 × 10¹⁷ | Likely | |
| 2008 | GNN | University of Siena | Robotics and other | 1.6 × 10⁹ | Confident |
| 2009 | LCNP MNIST | Vision, images and video | 4.2 × 10¹⁵ | Confident | |
| 2009 | LCNP LabelMe | University of Bonn | Vision, images and video | 3.3 × 10¹⁵ | Confident |
| 2009 | LCNP NORB | Vision, images and video | 2.5 × 10¹⁵ | Confident | |
| 2010 | RNN LM | Johns Hopkins University | Language | 5.4 × 10¹⁶ | Speculative |
| 2010 | Pooling CNN (NORB) | University of Bonn | Vision, images and video | 1.5 × 10¹⁵ | Confident |
| 2010 | Pooling CNN (Caltech 101) | University of Bonn | Vision, images and video | 1.2 × 10¹⁵ | Confident |
| 2011 | CNN Committee (MNIST) | IDSIA | Vision, images and video | 5.2 × 10¹⁶ | Likely |
| 2011 | Deep Autoencoders | University of Toronto | Vision, images and video | 3.7 × 10¹⁶ | Confident |
| 2011 | CNN Committee (NIST) | IDSIA | Vision, images and video | 2.6 × 10¹⁶ | Likely |
| 2012 | Unsupervised High-level Feature Learner | Vision, images and video | 6.0 × 10¹⁷ | Likely | |
| 2012 | DNN EM segmentation | IDSIA, SUPSI | Vision, images and video | 4.8 × 10¹⁷ | Confident |
| 2012 | AlexNet | University of Toronto | Vision, images and video | 4.7 × 10¹⁷ | Confident |
| 2013 | DistBelief NNLM | Language | 2.6 × 10¹⁸ | Likely | |
| 2013 | TransE | Universite de Technologie de Compiègne – CNRS, Google | Language | 1.3 × 10¹⁸ | Speculative |
| 2013 | Visualizing CNNs | New York University (NYU) | Vision, images and video | 5.3 × 10¹⁷ | Confident |
| 2014 | SNM-skip | Language | 3.0 × 10²⁰ | Speculative | |
| 2014 | Seq2Seq LSTM | Language | 5.6 × 10¹⁹ | Confident | |
| 2014 | VGG16 | University of Oxford | Vision, images and video | 1.2 × 10¹⁹ | Confident |
| 2015 | AlphaGo Fan | DeepMind | Games | 3.8 × 10²⁰ | Likely |
| 2015 | Inception v3 | Google, University College London (UCL) | Vision, images and video | 1.0 × 10²⁰ | Likely |
| 2015 | MSRA (C, PReLU) | Microsoft Research | Vision, images and video | 2.4 × 10¹⁹ | Confident |
| 2016 | GNMT | Language | 6.6 × 10²¹ | Likely | |
| 2016 | NASv3 (CIFAR-10) | Google Brain | Vision, images and video | 2.2 × 10²¹ | Likely |
| 2016 | AlphaGo Lee | DeepMind | Games | 1.9 × 10²¹ | Speculative |
| 2017 | JFT | Google Research, Carnegie Mellon University (CMU) | Vision, images and video | 8.4 × 10²⁰ | Confident |
| 2017 | AlphaGo Zero | DeepMind | Games | 6.5 × 10²⁰ | Confident |
| 2017 | OpenAI TI7 DOTA 1v1 | OpenAI | Games | 6.0 × 10²⁰ | Confident |
| 2018 | ResNeXt-101 32x48d | Vision, images and video | 8.7 × 10²¹ | Confident | |
| 2018 | Big Transformer for Back-Translation | Facebook AI Research, Google Brain | Language | 4.8 × 10²⁰ | Likely |
| 2018 | BERT-Large | Language | 2.9 × 10²⁰ | Confident | |
| 2019 | AlphaStar | DeepMind | Games | 1.1 × 10²³ | Confident |
| 2019 | OpenAI Five | OpenAI | Games | 6.7 × 10²² | Confident |
| 2019 | T5-11B | Language | 3.3 × 10²² | Confident | |
| 2020 | GPT-3 175B (davinci) | OpenAI | Language | 3.1 × 10²³ | Confident |
| 2020 | Meena | Google Brain | Language | 1.1 × 10²³ | Confident |
| 2020 | mT5-XXL | Google, Google Research | Language | 8.2 × 10²² | Confident |
| 2021 | FLAN 137B | Google Research | Language | 2.0 × 10²⁴ | Confident |
| 2021 | ERNIE 3.0 Titan | Baidu, Peng Cheng Laboratory | Language | 1.0 × 10²⁴ | Confident |
| 2021 | Megatron-Turing NLG 530B | Microsoft, Nvidia | Language | 8.6 × 10²³ | Confident |
| 2022 | Minerva (540B) | Language | 2.7 × 10²⁴ | Confident | |
| 2022 | GPT-3.5 (davinci-002) | OpenAI | Language | 2.6 × 10²⁴ | Speculative |
| 2022 | Flan-PaLM 540B | Language | 2.5 × 10²⁴ | Confident | |
| 2023 | Gemini 1.0 Ultra | Google DeepMind | Multimodal | 5.0 × 10²⁵ | Speculative |
| 2023 | GPT-4 (Jun 2023) | OpenAI | Multimodal | 2.1 × 10²⁵ | Likely |
| 2023 | GPT-4 (Mar 2023) | OpenAI | Multimodal | 2.1 × 10²⁵ | Likely |
| 2024 | Llama 3.1-405B | Meta AI | Language | 3.8 × 10²⁵ | Confident |
| 2024 | Grok-2 | xAI | Multimodal | 3.0 × 10²⁵ | Confident |
| 2024 | Claude 3.5 Sonnet | Anthropic | Multimodal | 2.7 × 10²⁵ | Speculative |
| 2025 | Grok 4 | xAI | Multimodal | 5.0 × 10²⁶ | Speculative |
| 2025 | GPT-4.5 | OpenAI | Multimodal | 3.8 × 10²⁶ | Likely |
| 2025 | Grok 3 | xAI | Multimodal | 3.5 × 10²⁶ | Likely |
| 2026 | GPT-6 Astra | OpenAI | Multimodal | 1.0 × 10²⁷ | Likely |
| 2026 | Composer 2.5 | Cursor | Language | 3.9 × 10²⁵ | Likely |
| 2026 | Composer 2 | Cursor | Language | 2.3 × 10²⁵ | Confident |
Source: Epoch AI, Data on AI models. notable_ai_models.csv · retrieved Sept. 26, 2026 · 540 rows · as of September 2026 · monthly. Download CSV
Benchmark scores Epoch AI ran itself
Every score here is Epoch AI's own evaluation, not a score reported by the company that built the model. Each chart shows one benchmark: a light point for each Epoch run, placed at the model's release date, and a step line for the best score to date. The line stops at the latest model Epoch has evaluated; it is not a projection.
Recent months are thinner than on Epoch's own site: a run is included only when Epoch's evaluation log for it is published, so 101 runs in Epoch's files for these benchmarks are not shown. As a rule of thumb this page uses, not a formal test, a step smaller than about two standard errors (given in each point's readout and the CSV) may be noise.
GPQA Diamond: best score to date
198 multiple-choice questions in biology, physics and chemistry, written by experts to be hard to answer by searching the web. Random guessing scores 25%.
Left and Right arrows move from point to point in date order; Page Up and Page Down jump a tenth of the way later or earlier; Home and End go to the first and last point; Escape clears.
Left and Right arrows move from point to point in date order; Page Up and Page Down jump a tenth of the way later or earlier; Home and End go to the first and last point; Escape clears.
Caveat. The expert line is the figure from the benchmark’s paper (81.3%), whose authors say validator accuracies on this set "are skewed by selection effects": the set keeps "only questions where both experts answer correctly and the majority of non-experts answer incorrectly." With 198 questions, a difference of a few points between two runs is noise. The benchmark is near saturation.
Data table
| Model released | Model | Organization | Best score to date | Previous best |
|---|---|---|---|---|
| 2026-02-19 | gemini-3.1-pro-preview_high | Google DeepMind | 94.4% | 92.6% (gemini-3-pro-preview) |
| 2025-11-18 | gemini-3-pro-preview | Google DeepMind | 92.6% | 87.6% (gpt-5.1-2025-11-13_high) |
| 2025-11-13 | gpt-5.1-2025-11-13_high | OpenAI | 87.6% | 86.2% (gpt-5-2025-08-07_high) |
| 2025-08-07 | gpt-5-2025-08-07_high | OpenAI | 86.2% | 85.3% (gemini-2.5-pro) |
| 2025-06-17 | gemini-2.5-pro | Google DeepMind | 85.3% | 84.8% (gemini-2.5-pro-preview-06-05) |
| 2025-06-05 | gemini-2.5-pro-preview-06-05 | Google DeepMind | 84.8% | 83.8% (gemini-2.5-pro-exp-03-25) |
| 2025-03-25 | gemini-2.5-pro-exp-03-25 | Google DeepMind | 83.8% | 79.7% (claude-3-7-sonnet-20250219_64K) |
| 2025-02-24 | claude-3-7-sonnet-20250219_64K | Anthropic | 79.7% | 77.0% (o3-mini-2025-01-31_high) |
| 2025-01-31 | o3-mini-2025-01-31_high | OpenAI | 77.0% | 76.8% (o1-2024-12-17_high) |
| 2024-12-17 | o1-2024-12-17_high | OpenAI | 76.8% | 62.4% (o1-mini-2024-09-12_high) |
| 2024-09-12 | o1-mini-2024-09-12_high | OpenAI | 62.4% | 54.0% (claude-3-5-sonnet-20240620) |
| 2024-06-20 | claude-3-5-sonnet-20240620 | Anthropic | 54.0% | 48.9% (gpt-4o-2024-05-13) |
| 2024-05-13 | gpt-4o-2024-05-13 | OpenAI | 48.9% | 47.2% (claude-3-opus-20240229) |
| 2024-02-29 | claude-3-opus-20240229 | Anthropic | 47.2% | 42.4% (gpt-4-1106-preview) |
| 2023-11-06 | gpt-4-1106-preview | OpenAI | 42.4% | 35.7% (gpt-4-0314) |
| 2023-03-14 | gpt-4-0314 | OpenAI | 35.7% | First run |
Source: Epoch AI, AI Benchmarking Hub. computed from epoch-benchmarks.csv · retrieved Sept. 26, 2026 · 93 rows · as of August 2026 · monthly. Download CSV
FrontierMath, Tiers 1 to 3: best score to date
The benchmark's 2024 paper describes hundreds of original, unpublished math problems written and checked by mathematicians, a typical one taking a researcher in the field hours. No human score is reported. The chart shows Epoch's second version of this set, "FrontierMath-Tiers-1-3-v2-Private", which Epoch's metadata dates June 12, 2026: newer than the 2024 paper, which describes the original set.
Left and Right arrows move from point to point in date order; Page Up and Page Down jump a tenth of the way later or earlier; Home and End go to the first and last point; Escape clears.
Left and Right arrows move from point to point in date order; Page Up and Page Down jump a tenth of the way later or earlier; Home and End go to the first and last point; Escape clears.
Caveat. OpenAI commissioned and funded FrontierMath. In January 2025, Epoch said OpenAI had access to most of the problems in the original set and their solutions, apart from a holdout set. The problems are private, so outside researchers cannot check them.
Data table
| Model released | Model | Organization | Best score to date | Previous best |
|---|---|---|---|---|
| 2026-07-09 | gpt-5.6-sol_max | OpenAI | 89.1% | 87.0% (claude-fable-5_max) |
| 2026-06-09 | claude-fable-5_max | Anthropic | 87.0% | 82.5% (gpt-5.4-pro-2026-03-05_xhigh) |
| 2026-03-05 | gpt-5.4-pro-2026-03-05_xhigh | OpenAI | 82.5% | 66.0% (claude-opus-4-6_max) |
| 2026-02-05 | claude-opus-4-6_max | Anthropic | 66.0% | 55.8% (gpt-5-pro-2025-10-06_high) |
| 2025-10-07 | gpt-5-pro-2025-10-06_high | OpenAI | 55.8% | 46.7% (gpt-5-mini-2025-08-07_high) |
| 2025-08-07 | gpt-5-mini-2025-08-07_high | OpenAI | 46.7% | 36.1% (o4-mini-2025-04-16_high) |
| 2025-04-16 | o4-mini-2025-04-16_high | OpenAI | 36.1% | 18.6% (o3-mini-2025-01-31_high) |
| 2025-01-31 | o3-mini-2025-01-31_high | OpenAI | 18.6% | 14.7% (o1-2024-12-17_high) |
| 2024-12-17 | o1-2024-12-17_high | OpenAI | 14.7% | 0.7% (gpt-4-turbo-2024-04-09) |
| 2024-04-09 | gpt-4-turbo-2024-04-09 | OpenAI | 0.7% | 0.0% (gpt-3.5-turbo-0125) |
| 2024-01-25 | gpt-3.5-turbo-0125 | OpenAI | 0.0% | First run |
Source: Epoch AI, AI Benchmarking Hub. computed from epoch-benchmarks.csv · retrieved Sept. 26, 2026 · 93 rows · as of August 2026 · monthly. Download CSV
FrontierMath, Tier 4: best score to date
FrontierMath’s Tier 4, scored separately from Tiers 1 to 3. No human score is reported. The chart shows Epoch's second version of this set, "FrontierMath-Tier-4-v2-Private", which Epoch's metadata dates June 12, 2026: newer than the 2024 paper, which describes the original set.
Left and Right arrows move from point to point in date order; Page Up and Page Down jump a tenth of the way later or earlier; Home and End go to the first and last point; Escape clears.
Left and Right arrows move from point to point in date order; Page Up and Page Down jump a tenth of the way later or earlier; Home and End go to the first and last point; Escape clears.
Caveat. As for Tiers 1 to 3: OpenAI commissioned and funded FrontierMath. In January 2025, Epoch said OpenAI had access to most of the problems in the original set and their solutions, apart from a holdout set. The problems are private, so outside researchers cannot check them.
Data table
| Model released | Model | Organization | Best score to date | Previous best |
|---|---|---|---|---|
| 2026-06-09 | claude-fable-5_max | Anthropic | 90.2% | 58.5% (gpt-5.4-pro-2026-03-05_xhigh) |
| 2026-03-05 | gpt-5.4-pro-2026-03-05_xhigh | OpenAI | 58.5% | 26.8% (claude-opus-4-6_max) |
| 2026-02-05 | claude-opus-4-6_max | Anthropic | 26.8% | 19.5% (gpt-5-pro-2025-10-06_high) |
| 2025-10-07 | gpt-5-pro-2025-10-06_high | OpenAI | 19.5% | 12.2% (gpt-5-mini-2025-08-07_high) |
| 2025-08-07 | gpt-5-mini-2025-08-07_high | OpenAI | 12.2% | 4.9% (o4-mini-2025-04-16_high) |
| 2025-04-16 | o4-mini-2025-04-16_high | OpenAI | 4.9% | 0.0% (o3-mini-2025-01-31_high) |
| 2025-01-31 | o3-mini-2025-01-31_high | OpenAI | 0.0% | First run |
Source: Epoch AI, AI Benchmarking Hub. computed from epoch-benchmarks.csv · retrieved Sept. 26, 2026 · 93 rows · as of August 2026 · monthly. Download CSV
SWE-bench Verified: best score to date
Real issues from GitHub projects written in Python: the model edits the code, and the project’s tests decide whether the fix works. Epoch runs 484 of the 500 tasks. No human baseline is reported.
Left and Right arrows move from point to point in date order; Page Up and Page Down jump a tenth of the way later or earlier; Home and End go to the first and last point; Escape clears.
Left and Right arrows move from point to point in date order; Page Up and Page Down jump a tenth of the way later or earlier; Home and End go to the first and last point; Escape clears.
Caveat. Epoch notes that some tasks may remain ambiguous and has estimated an error rate of 5 to 10%. OpenAI, which published SWE-bench Verified, said in February 2026 that it had stopped reporting scores on it: in an audit of 138 problems its o3 model did not consistently solve, it found that 59.4% had material issues in test design or problem description, and it reported signs that every frontier model it tested had seen at least some of the problems and solutions during training. In a separate METR study of an AI agent’s pull requests for tasks from two open-source projects, not SWE-bench, none of the 15 that METR reviewed by hand was mergeable as-is, including those that passed the projects’ own tests.
Data table
| Model released | Model | Organization | Best score to date | Previous best |
|---|---|---|---|---|
| 2026-04-16 | claude-opus-4-7_max | Anthropic | 83.5% | 78.7% (claude-opus-4-6) |
| 2026-02-05 | claude-opus-4-6 | Anthropic | 78.7% | 76.7% (claude-opus-4-5-20251101) |
| 2025-11-24 | claude-opus-4-5-20251101 | Anthropic | 76.7% | 73.6% (gpt-5-2025-08-07_high) |
| 2025-08-07 | gpt-5-2025-08-07_high | OpenAI | 73.6% | 73.3% (claude-opus-4-1-20250805) |
| 2025-08-05 | claude-opus-4-1-20250805 | Anthropic | 73.3% | 70.7% (claude-opus-4-20250514) |
| 2025-05-22 | claude-opus-4-20250514 | Anthropic | 70.7% | 62.3% (o3-2025-04-16_medium) |
| 2025-04-16 | o3-2025-04-16_medium | OpenAI | 62.3% | 61.0% (claude-3-7-sonnet-20250219) |
| 2025-02-24 | claude-3-7-sonnet-20250219 | Anthropic | 61.0% | 31.0% (gpt-4o-2024-11-20) |
| 2024-11-20 | gpt-4o-2024-11-20 | OpenAI | 31.0% | First run |
Source: Epoch AI, AI Benchmarking Hub. computed from epoch-benchmarks.csv · retrieved Sept. 26, 2026 · 93 rows · as of August 2026 · monthly. Download CSV
MATH Level 5: best score to date
A benchmark Epoch AI runs itself, listed as "MATH level 5" on Epoch's benchmarking hub (epoch.ai/benchmarks).
Left and Right arrows move from point to point in date order; Page Up and Page Down jump a tenth of the way later or earlier; Home and End go to the first and last point; Escape clears.
Left and Right arrows move from point to point in date order; Page Up and Page Down jump a tenth of the way later or earlier; Home and End go to the first and last point; Escape clears.
Caveat. Top scores are near the ceiling: shown for its history, not to rank current models.
Data table
| Model released | Model | Organization | Best score to date | Previous best |
|---|---|---|---|---|
| 2025-08-07 | gpt-5-2025-08-07_high | OpenAI | 98.1% | 97.8% (o4-mini-2025-04-16_high) |
| 2025-04-16 | o4-mini-2025-04-16_high | OpenAI | 97.8% | 96.5% (o3-mini-2025-01-31_high) |
| 2025-01-31 | o3-mini-2025-01-31_high | OpenAI | 96.5% | 94.7% (o1-2024-12-17_high) |
| 2024-12-17 | o1-2024-12-17_high | OpenAI | 94.7% | 89.2% (o1-mini-2024-09-12_high) |
| 2024-09-12 | o1-mini-2024-09-12_high | OpenAI | 89.2% | 53.3% (gpt-4o-2024-08-06) |
| 2024-08-06 | gpt-4o-2024-08-06 | OpenAI | 53.3% | 52.6% (gpt-4o-mini-2024-07-18) |
| 2024-07-18 | gpt-4o-mini-2024-07-18 | OpenAI | 52.6% | 51.7% (claude-3-5-sonnet-20240620) |
| 2024-06-20 | claude-3-5-sonnet-20240620 | Anthropic | 51.7% | 51.0% (gpt-4o-2024-05-13) |
| 2024-05-13 | gpt-4o-2024-05-13 | OpenAI | 51.0% | 46.7% (gpt-4-turbo-2024-04-09) |
| 2024-04-09 | gpt-4-turbo-2024-04-09 | OpenAI | 46.7% | 40.0% (gpt-4-1106-preview) |
| 2023-11-06 | gpt-4-1106-preview | OpenAI | 40.0% | 23.0% (gpt-4-0613) |
| 2023-06-13 | gpt-4-0613 | OpenAI | 23.0% | First run |
Source: Epoch AI, AI Benchmarking Hub. computed from epoch-benchmarks.csv · retrieved Sept. 26, 2026 · 93 rows · as of August 2026 · monthly. Download CSV
OTIS Mock AIME 2024-2025: best score to date
A benchmark Epoch AI runs itself, listed as "OTIS Mock AIME 2024-2025" on Epoch's benchmarking hub (epoch.ai/benchmarks).
Left and Right arrows move from point to point in date order; Page Up and Page Down jump a tenth of the way later or earlier; Home and End go to the first and last point; Escape clears.
Left and Right arrows move from point to point in date order; Page Up and Page Down jump a tenth of the way later or earlier; Home and End go to the first and last point; Escape clears.
Caveat. Several models score 100% in Epoch’s runs: shown for its history, not to rank current models.
Data table
| Model released | Model | Organization | Best score to date | Previous best |
|---|---|---|---|---|
| 2026-04-23 | gpt-5.5-pre-release_xhigh | OpenAI | 100.0% | 97.8% (claude-opus-4-7_xhigh) |
| 2026-04-16 | claude-opus-4-7_xhigh | Anthropic | 97.8% | 97.8% (gpt-5.4-2026-03-05_high) |
| 2026-03-05 | gpt-5.4-2026-03-05_high | OpenAI | 97.8% | 96.1% (gpt-5.2-2025-12-11_high) |
| 2025-12-11 | gpt-5.2-2025-12-11_high | OpenAI | 96.1% | 91.4% (gpt-5-2025-08-07_high) |
| 2025-08-07 | gpt-5-2025-08-07_high | OpenAI | 91.4% | 88.9% (openai/gpt-oss-120b_high) |
| 2025-08-05 | openai/gpt-oss-120b_high | OpenAI | 88.9% | 86.7% (Qwen3-235B-A22B-Thinking-2507) |
| 2025-07-25 | Qwen3-235B-A22B-Thinking-2507 | Alibaba | 86.7% | 84.7% (gemini-2.5-pro) |
| 2025-06-17 | gemini-2.5-pro | Google DeepMind | 84.7% | 84.4% (o3-2025-04-16_medium) |
| 2025-04-16 | o3-2025-04-16_medium | OpenAI | 84.4% | 77.8% (grok-3-mini-beta_high) |
| 2025-04-09 | grok-3-mini-beta_high | xAI | 77.8% | 76.9% (o3-mini-2025-01-31_high) |
| 2025-01-31 | o3-mini-2025-01-31_high | OpenAI | 76.9% | 73.3% (o1-2024-12-17_high) |
| 2024-12-17 | o1-2024-12-17_high | OpenAI | 73.3% | 46.9% (o1-mini-2024-09-12_high) |
| 2024-09-12 | o1-mini-2024-09-12_high | OpenAI | 46.9% | 9.7% (Llama-3.1-405B-Instruct) |
| 2024-07-23 | Llama-3.1-405B-Instruct | Meta AI | 9.7% | 6.9% (gpt-4o-mini-2024-07-18) |
| 2024-07-18 | gpt-4o-mini-2024-07-18 | OpenAI | 6.9% | 6.8% (gemini-1.5-pro-001) |
| 2024-05-14 | gemini-1.5-pro-001 | Google DeepMind | 6.8% | 6.7% (gpt-4-turbo-2024-04-09) |
| 2024-04-09 | gpt-4-turbo-2024-04-09 | OpenAI | 6.7% | 4.7% (claude-3-opus-20240229) |
| 2024-02-29 | claude-3-opus-20240229 | Anthropic | 4.7% | 2.5% (claude-2.0) |
| 2023-07-11 | claude-2.0 | Anthropic | 2.5% | 1.1% (gpt-4-0613) |
| 2023-06-13 | gpt-4-0613 | OpenAI | 1.1% | 0.6% (gpt-4-0314) |
| 2023-03-14 | gpt-4-0314 | OpenAI | 0.6% | First run |
Source: Epoch AI, AI Benchmarking Hub. computed from epoch-benchmarks.csv · retrieved Sept. 26, 2026 · 93 rows · as of August 2026 · monthly. Download CSV
Share of U.S. businesses using AI
The Census Bureau's Business Trends and Outlook Survey asks businesses every two weeks whether they used AI in the previous two weeks. In November 2025 the Census Bureau changed the question's wording and, because of the shift in answers that came with it, started a new series. The two lines are separate series and are not comparable; the chart does not join them.
Share of U.S. businesses that used AI in the last two weeks
Percent of responding businesses answering yes, by two-week reference period. The shaded band marks 3 periods with no estimate: the survey was not collected during the federal funding lapse. BTOS is an experimental data product.
Left and Right arrows move from point to point in date order; Page Up and Page Down jump a tenth of the way later or earlier; Home and End go to the first and last point; Escape clears.
Left and Right arrows move from point to point in date order; Page Up and Page Down jump a tenth of the way later or earlier; Home and End go to the first and last point; Escape clears.
Where the rule sits. Each point is placed at the last day of the two weeks it describes. The “New question” rule is at the two weeks ending Nov. 16, 2025 (Nov. 3 to 16, 2025), the first period answered with the new wording. Businesses were first asked it on Nov. 17, 2025, when collection for that period began, so the rule sits a day before the wording changed.
- Original question, Aug. 28, 2023 to Sept. 21, 2025 (reference periods)
In the last two weeks, did this business use Artificial Intelligence (AI) in producing goods or services? (Examples of AI: machine learning, natural language processing, virtual agents, voice recognition, etc.)
- Updated question, asked from Nov. 17, 2025
In the last two weeks, did this business use Artificial Intelligence (AI) in any of its business functions? (Examples of AI: machine learning, natural language processing, virtual agents, voice recognition, etc.)
Why two series. The Census Bureau: Due to a level shift observed in conjunction with the new question wording, the decision was made to create a new time series for the AI questions, beginning with data released on December 4, 2025.
In the Census Bureau's testing, some respondents did not see their business as a producer of goods or services, and others answered no while using AI in work such as hiring, accounting or software development.
Data table
| Two weeks ending | Period began | Original question | Updated question | Note |
|---|---|---|---|---|
| 2026-09-06 | 2026-08-24 | 23.8% | ||
| 2026-08-23 | 2026-08-10 | 23.2% | ||
| 2026-08-09 | 2026-07-27 | 22.4% | ||
| 2026-07-26 | 2026-07-13 | 21.8% | ||
| 2026-07-12 | 2026-06-29 | 21.5% | ||
| 2026-06-28 | 2026-06-15 | 21.7% | ||
| 2026-06-14 | 2026-06-01 | 20.6% | ||
| 2026-05-31 | 2026-05-18 | 20.6% | ||
| 2026-05-17 | 2026-05-04 | 20.6% | ||
| 2026-05-03 | 2026-04-20 | 19.5% | ||
| 2026-04-19 | 2026-04-06 | 19.8% | ||
| 2026-04-05 | 2026-03-23 | 19.8% | ||
| 2026-03-22 | 2026-03-09 | 19.1% | ||
| 2026-03-08 | 2026-02-23 | 18.9% | ||
| 2026-02-22 | 2026-02-09 | 18.2% | ||
| 2026-02-08 | 2026-01-26 | 17.5% | ||
| 2026-01-25 | 2026-01-12 | 18.9% | ||
| 2026-01-11 | 2025-12-29 | 18.2% | ||
| 2025-12-28 | 2025-12-15 | 17.7% | ||
| 2025-12-14 | 2025-12-01 | 17.8% | ||
| 2025-11-30 | 2025-11-17 | 17.2% | ||
| 2025-11-16 | 2025-11-03 | 17.3% | ||
| 2025-11-02 | 2025-10-20 | No estimate: BTOS did not collect data for this period (federal funding lapse, October 1 to November 13, 2025) | ||
| 2025-10-19 | 2025-10-06 | No estimate: BTOS did not collect data for this period (federal funding lapse, October 1 to November 13, 2025) | ||
| 2025-10-05 | 2025-09-22 | No estimate: BTOS did not collect data for this period (federal funding lapse, October 1 to November 13, 2025) | ||
| 2025-09-21 | 2025-09-08 | 10.0% | ||
| 2025-09-07 | 2025-08-25 | 9.9% | ||
| 2025-08-24 | 2025-08-11 | 9.1% | ||
| 2025-08-10 | 2025-07-28 | 9.7% | ||
| 2025-07-27 | 2025-07-14 | 8.8% | ||
| 2025-07-13 | 2025-06-30 | 9.4% | ||
| 2025-06-29 | 2025-06-16 | 9.3% | ||
| 2025-06-15 | 2025-06-02 | 9.3% | ||
| 2025-06-01 | 2025-05-19 | 9.2% | ||
| 2025-05-18 | 2025-05-05 | 9.2% | ||
| 2025-05-04 | 2025-04-21 | 8.7% | ||
| 2025-04-20 | 2025-04-07 | 8.7% | ||
| 2025-04-06 | 2025-03-24 | 8.3% | ||
| 2025-03-23 | 2025-03-10 | 8.0% | ||
| 2025-03-09 | 2025-02-24 | 7.8% | ||
| 2025-02-23 | 2025-02-10 | 7.4% | ||
| 2025-02-09 | 2025-01-27 | 7.3% | ||
| 2025-01-26 | 2025-01-13 | 6.8% | ||
| 2025-01-12 | 2024-12-30 | 6.7% | ||
| 2024-12-29 | 2024-12-16 | 6.0% | ||
| 2024-12-15 | 2024-12-02 | 6.3% | ||
| 2024-12-01 | 2024-11-18 | 6.1% | ||
| 2024-11-17 | 2024-11-04 | 6.1% | ||
| 2024-11-03 | 2024-10-21 | 6.1% | ||
| 2024-10-20 | 2024-10-07 | 6.0% | ||
| 2024-10-06 | 2024-09-23 | 6.0% | ||
| 2024-09-22 | 2024-09-09 | 6.0% | ||
| 2024-09-08 | 2024-08-26 | 5.9% | ||
| 2024-08-25 | 2024-08-12 | 5.7% | ||
| 2024-08-11 | 2024-07-29 | 5.5% | ||
| 2024-07-28 | 2024-07-15 | 5.1% | ||
| 2024-07-14 | 2024-07-01 | 5.1% | ||
| 2024-06-30 | 2024-06-17 | 5.0% | ||
| 2024-06-16 | 2024-06-03 | 4.7% | ||
| 2024-06-02 | 2024-05-20 | 5.0% | ||
| 2024-05-19 | 2024-05-06 | 4.6% | ||
| 2024-05-05 | 2024-04-22 | 4.8% | ||
| 2024-04-21 | 2024-04-08 | 4.7% | ||
| 2024-04-07 | 2024-03-25 | 4.8% | ||
| 2024-03-24 | 2024-03-11 | 4.2% | ||
| 2024-03-10 | 2024-02-26 | 4.4% | ||
| 2024-02-25 | 2024-02-12 | 4.3% | ||
| 2024-02-11 | 2024-01-29 | 5.4% | ||
| 2024-01-28 | 2024-01-15 | 5.0% | ||
| 2024-01-14 | 2024-01-01 | 5.3% | ||
| 2023-12-31 | 2023-12-18 | 4.6% | ||
| 2023-12-17 | 2023-12-04 | 4.9% | ||
| 2023-12-03 | 2023-11-20 | 4.9% | ||
| 2023-11-19 | 2023-11-06 | 4.6% | ||
| 2023-11-05 | 2023-10-23 | 4.4% | ||
| 2023-10-22 | 2023-10-09 | 3.9% | ||
| 2023-10-08 | 2023-09-25 | 3.8% | ||
| 2023-09-24 | 2023-09-11 | 3.7% | ||
| 2023-09-10 | 2023-08-28 | 3.7% |
Source: U.S. Census Bureau, Business Trends and Outlook Survey (BTOS). BTOS question 7, answer Yes, national: National.xlsx (updated wording) and AI Core Questions.xlsx, sheet National Estimates (original wording); standard errors from each file's SE sheet · retrieved Sept. 26, 2026 · 79 rows · as of Sept. 6, 2026 · biweekly. Question wording: BTOS AI Core Question Updates, December 3, 2025. Download CSV
Milestones
Dated events, each with the primary source read for it. The second line of each entry is limited to what that source states. The International Mathematical Olympiad results of 2025 are marked disputed, with each view attributed.
- IBM's Deep Blue beat world chess champion Garry Kasparov 3.5 to 2.5 in a six-game rematch in New York.A computer won a match under standard time controls against a reigning world champion.Source: Deep Blue
- The SuperVision team (Krizhevsky, Sutskever and Hinton; the network known as AlexNet) won the ILSVRC 2012 ImageNet classification task.Its top-5 error was 0.15315 (with extra ImageNet Fall 2011 data), against 0.26172 for the next team (ISI).Source: ILSVRC2012 results
- DeepMind posted "Playing Atari with Deep Reinforcement Learning" (DQN).The authors describe "the first deep learning model to successfully learn control policies directly from high-dimensional sensory input using reinforcement learning," tested on seven Atari games and better than a human expert on three.Source: Playing Atari with Deep Reinforcement Learning
- DeepMind's AlphaGo beat Fan Hui 5 to 0 (October 2015), then Lee Sedol 4 to 1 in Seoul (March 2016).DeepMind calls the Fan Hui games "the first ever match between an AI system and Go professional," and AlphaGo received a 9 dan professional ranking.Source: AlphaGo
- Vaswani and colleagues at Google posted "Attention Is All You Need".The paper proposed "the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely."Source: Attention Is All You Need
- Devlin and colleagues at Google posted BERT.Pre-trained bidirectional representations reached a GLUE score of 80.5% ("7.7% point absolute improvement").Source: BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
- Brown and colleagues at OpenAI posted the GPT-3 paper.The paper describes "an autoregressive language model with 175 billion parameters," tested in the few-shot setting.Source: Language Models are Few-Shot Learners
- AlphaFold 2 results at CASP14.The CASP press release says AlphaFold determined "the shape of around two thirds of the proteins with accuracy comparable to laboratory experiments," for "single proteins or domains, not protein complexes."Source: CASP14 press release
- Wei and colleagues at Google posted the chain-of-thought prompting paper.The paper reports that "generating a chain of thought -- a series of intermediate reasoning steps -- significantly improves the ability of large language models to perform complex reasoning."Source: Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
- DeepSeek posted DeepSeek-R1, an open-weight reasoning model.Reasoning was trained through reinforcement learning; the work was published in Nature 645 (2025).Source: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
- DisputedAI systems' results on the 2025 International Mathematical Olympiad problems.Google DeepMind reported 35 of 42 points for an advanced version of Gemini Deep Think, with solutions graded and certified by IMO coordinators (Google DeepMind, July 21, 2025). OpenAI published proofs produced by its experimental reasoning model on the 2025 problems. Google DeepMind's post says the certification covers the answers, not the system. Per mathematician Kevin Buzzard's round-up (August 3, 2025), OpenAI's solutions were not graded by IMO coordinators; OpenAI's own announcement was not read for this page, so no OpenAI score is given here. The IMO president said the IMO "cannot validate the methods, including the amount of compute used or whether there was any human involvement, or whether the results can be reproduced. What we can say is that correct mathematical proofs, whether produced by the brightest students or AI models, are valid."Source: The 66th International Mathematical Olympiad draws to a close todayAlso: Google DeepMind's post (July 21, 2025), Kevin Buzzard's round-up (August 3, 2025), OpenAI's published proofs
3 more entries, whose dates could not yet be confirmed from a primary source, are in the milestones file but not shown. The milestones CSV lists each with the reason.
Sources and methods
Epoch AI
The training compute and benchmark data are from Epoch AI and are used under the Creative Commons Attribution 4.0 license. Epoch's attribution:
Epoch AI, 'Data on AI models'. Published online at epoch.ai. Retrieved from 'https://epoch.ai/data/ai-models' [online resource].
Epoch AI, 'Capabilities & benchmarking'. Published online at epoch.ai. Retrieved from 'https://epoch.ai/benchmarks/use-this-data'.
Benchmark runs are kept only when their evaluation log is in Epoch's own evaluation records, so every score on this page is one Epoch ran. The best score to date is computed here from those runs, per benchmark and by model release date; the superseded versions of the FrontierMath sets are left out. Compute figures are as Epoch publishes them, including its estimates; nothing is rounded, filled or fitted, and no trend line is drawn.
SWE-bench Verified caveat
- Epoch AI, SWE-bench Verified, accessed Sept. 26, 2026, archived copy, Sept. 18, 2026.
- OpenAI, Why SWE-bench Verified no longer measures frontier coding capabilities (Feb. 23, 2026), accessed Sept. 26, 2026, archived copy, Sept. 15, 2026.
- METR, Research Update: Algorithmic vs. Holistic Evaluation (Aug. 13, 2025), accessed Sept. 26, 2026, archived copy, Sept. 20, 2026.
U.S. Census Bureau
Business Trends and Outlook Survey: the updated question from National.xlsx, the original question from AI Core Questions.xlsx, and the wording change from BTOS AI Core Question Updates (December 3, 2025).
What is not here, and why
METR's task time-horizon measurements are left out until METR grants permission to republish them; no license covers that data. ARC Prize's verified ARC-AGI scores are left out because the terms bar commercial republishing without written permission. Humanity's Last Exam is left out because its leaderboard does not state who ran each evaluation and no license for its results was found; FutureHouse, an independent research lab, also estimated from a checked subset that about 29% of its text-only chemistry and biology answers conflict with evidence in peer-reviewed literature. MMLU is left out because it is saturated (the Humanity's Last Exam paper says "LLMs now achieve over 90% accuracy on popular benchmarks like MMLU"), and "Are We Done with MMLU?" (Gema and colleagues) estimated that 6.49% of its questions contain errors; no license for published MMLU results was found. The Stanford AI Index compiles results rather than running tests (it "assumes that company-reported results are accurate"), and the license on its arXiv listing (CC BY-NC-ND 4.0) bars adapted charts. Scores that companies report for their own models are not used anywhere on this page.
The data
- Training compute of notable AI models (Epoch AI): CSV of Training compute of notable AI models (Epoch AI), catalog entry for Training compute of notable AI models (Epoch AI)
- AI benchmark results run by Epoch AI: CSV of AI benchmark results run by Epoch AI, catalog entry for AI benchmark results run by Epoch AI
- Best Epoch-run benchmark score to date, by model release date: CSV of Best Epoch-run benchmark score to date, by model release date, catalog entry for Best Epoch-run benchmark score to date, by model release date
- Share of U.S. businesses using AI (Census BTOS), two question versions: CSV of Share of U.S. businesses using AI (Census BTOS), two question versions, catalog entry for Share of U.S. businesses using AI (Census BTOS), two question versions
- AI milestones with their primary sources: CSV of AI milestones with their primary sources, catalog entry for AI milestones with their primary sources
This site uses AI tools; see the note on our About page.