AI Safety Tracker

AI training compute and cost

540 models with compute 1083 tracked 6 legal thresholds

Training compute, disclosed cost and release decisions for 540 notable AI models from Epoch AI, and the 6 training-compute thresholds written into law or policy.

How fast is AI improving?

Doubling time of frontier training compute Fitted by this site to Epoch's compute figures, not a figure Epoch publishes. Only the models that set a new record enter the fit, which is a small number of points. Starting the window at 2012, 2018 or 2020 instead gives 5.2, 5.0 and 5.2 months, so the answer is not an artefact of where the window starts. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , today Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/capability/.

6months

Least squares on the largest run to date, 2015 onward. A description of what happened, not a projection.

Largest reported training run Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , today Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/capability/.

9.4e26FLOP

GPT-6 Astra, OpenAI, 2026-09-03.

Models released with open weights in 2025 Accessibility is unrecorded for a substantial minority; those are counted separately, not as unreleased. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , today Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/capability/.

42%

Of 107 notable models published that year. Weights are the one release decision nobody can reverse.

Models at or above the EU systemic-risk compute threshold Counted from models with a reported or estimated compute figure. Many values are estimates. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , today Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/capability/.

24

Article 51 of the AI Act sets it at 1e25 FLOP, the lowest of the three compute thresholds currently in force.

Models at or above 1e26 FLOP, the Californian threshold Counted by this site from Epoch's compute figures against the threshold written in the instrument. Not a legal finding about any model. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , 5 days ago Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 23 October 2026 via AI Safety Tracker, https://aisafetytracker.org/capability/. Compute Threshold Index This site Licence MIT Updated quarterly Retrieved , 5 days ago Data JSON CSV Training-compute thresholds written into law, executive action or published policy, with the text that sets each one. Whether a given model crosses a threshold here is arithmetic on Epoch's compute estimates, most of which are estimates rather than disclosures, and is not a legal finding about any model or developer. Cite AI Safety Tracker, "Compute Threshold Index". https://aisafetytracker.org/capability/#thresholds. Retrieved 23 October 2026.

4

SB 53 applies to frontier developers above this line. 3 compute thresholds are in force anywhere in the world.

  • 1e25 24
  • 1e26 4
  • 1e26 4
  • 1e26 (revoked) 4
  • 1e26 (vetoed) 4
  • 1e26 (enacted) 4

Training compute over time

Every notable model with a reported or estimated training-compute figure, on a logarithmic scale. Each step of the axis is ten times the one below it.

1e141e161e181e201e221e241e26201220142016201820202022202420261e25 FLOP, EU AI Act Article 51Dropout (MNIST) (University of Toronto, 2012-06-03): 6.0e15 FLOP, confidence ConfidentDropout (ImageNet) (University of Toronto, 2012-06-03): 2.7e17 FLOP, confidence ConfidentDropout (CIFAR) (University of Toronto, 2012-06-03): 4.3e15 FLOP, confidence ConfidentUnsupervised High-level Feature Learner (Google, 2012-07-12): 6.0e17 FLOP, confidence LikelyLSTM LM (RWTH Aachen University, 2012-09-09): 1.7e16 FLOP, confidence SpeculativeAlexNet (University of Toronto, 2012-09-30): 4.7e17 FLOP, confidence ConfidentDNN EM segmentation (IDSIA,SUPSI, 2012-12-03): 4.8e17 FLOP, confidence ConfidentDistBelief Speech (Google, 2012-12-03): 3.1e17 FLOP, confidence SpeculativeDistBelief NNLM (Google, 2013-01-16): 2.6e18 FLOP, confidence LikelyReLU-Speech (Google,University of Toronto,New York University (NYU), 2013-05-26): 1.3e17 FLOP, confidence LikelyHierarchical Scene Labeling (Stanford Background) (New York University (NYU), 2013-08-01): 2.4e17 FLOP, confidence ConfidentRNTN (Stanford University, 2013-10-01): 1.4e16 FLOP, confidence LikelyRCTM (University of Oxford, 2013-10-01): 9.3e15 FLOP, confidence LikelyWord2Vec (large) (Google, 2013-10-16): 3.9e16 FLOP, confidence LikelyVisualizing CNNs (New York University (NYU), 2013-11-12): 5.3e17 FLOP, confidence ConfidentTransE (Universite de Technologie de Compiègne - CNRS,Google, 2013-12-05): 1.3e18 FLOP, confidence SpeculativeDQN (DeepMind, 2013-12-19): 2.8e15 FLOP, confidence ConfidentImage generation (University of Amsterdam, 2013-12-20): 4.8e14 FLOP, confidence ConfidentGANs (University of Montreal / Université de Montréal, 2014-06-10): 5.2e17 FLOP, confidence SpeculativeSPPNet (Microsoft,Xi’an Jiaotong University,University of Science and Technology of China (USTC), 2014-06-18): 3.4e18 FLOP, confidence ConfidentSmooCT (University College London (UCL), 2014-07-01): 6.9e16 FLOP, confidence LikelyACF-WIDER (Chinese Academy of Sciences, 2014-07-15): 7.6e13 FLOP, confidence ConfidentRNNsearch-50* (Jacobs University Bremen,University of Montreal / Université de Montréal, 2014-09-01): 1.6e18 FLOP, confidence ConfidentVGG19 (University of Oxford, 2014-09-04): 1.1e19 FLOP, confidence LikelyVGG16 (University of Oxford, 2014-09-04): 1.2e19 FLOP, confidence ConfidentSeq2Seq LSTM (Google, 2014-09-10): 5.6e19 FLOP, confidence ConfidentSPN-4+KN5 (Singapore University of Technology & Design,DSO National Laboratories, 2014-09-14): 4.4e16 FLOP, confidence LikelyGoogLeNet / InceptionV1 (Google,University of Michigan,University of North Carolina, 2014-09-17): 1.5e18 FLOP, confidence ConfidentTA-CNN (Chinese University of Hong Kong (CUHK), 2014-11-29): 1.1e16 FLOP, confidence LikelySNM-skip (Google, 2014-12-03): 3.0e20 FLOP, confidence SpeculativeFractional Max-Pooling (University of Warwick, 2014-12-18): 1.0e17 FLOP, confidence LikelyADAM (CIFAR-10) (University of Amsterdam,OpenAI,University of Toronto, 2014-12-22): 6.2e14 FLOP, confidence ConfidentMSRA (C, PReLU) (Microsoft Research, 2015-02-06): 2.4e19 FLOP, confidence ConfidentgenCNN + dyn eval (Chinese Academy of Sciences,Huawei Noah's Ark Lab,Dublin City University, 2015-03-17): 3.4e16 FLOP, confidence SpeculativeTC-DNN-BLSTM-DNN (Carnegie Mellon University (CMU), 2015-04-06): 1.9e17 FLOP, confidence ConfidentU-Net (University of Freiburg, 2015-05-18): 5.1e16 FLOP, confidence ConfidentDCNN (University of Maryland,Rutgers University, 2015-08-07): 4.8e17 FLOP, confidence ConfidentAlphaGo Fan (DeepMind, 2015-10-01): 3.8e20 FLOP, confidence LikelySAF R-CNN (Beijing Institute of Technology,Sun Yat-sen University,Panasonic R&D,National University of Singapore, 2015-10-28): 1.2e19 FLOP, confidence LikelyInception v3 (Google,University College London (UCL), 2015-12-02): 1.0e20 FLOP, confidence LikelyResNet-152 (ImageNet) (Microsoft, 2015-12-10): 1.0e19 FLOP, confidence ConfidentResNet-101 (ImageNet) (Microsoft, 2015-12-10): 7.0e18 FLOP, confidence ConfidentVariational (untied weights, MC) LSTM (Large) (University of Cambridge, 2015-12-16): 5.9e15 FLOP, confidence ConfidentAlphaGo Lee (DeepMind, 2016-01-27): 1.9e21 FLOP, confidence SpeculativeNamed Entity Recognition model (Carnegie Mellon University (CMU), 2016-03-04): 9.7e16 FLOP, confidence ConfidentR-FCN (Tsinghua University,Microsoft Research, 2016-06-21): 7.2e17 FLOP, confidence ConfidentResNet-200 (Microsoft Research Asia, 2016-09-17): 3.0e19 FLOP, confidence SpeculativePointer Sentinel-LSTM (medium) (MetaMind Inc,Salesforce, 2016-09-26): 7.5e15 FLOP, confidence ConfidentGNMT (Google, 2016-09-26): 6.6e21 FLOP, confidence LikelyXception (Google, 2016-10-07): 4.4e20 FLOP, confidence ConfidentSPIDER2 (Griffith University,University of Iowa,Dezhou University, 2016-10-28): 1.8e16 FLOP, confidence LikelyVD-LSTM+REAL Large (Salesforce Research,Stanford University, 2016-11-04): 2.1e16 FLOP, confidence ConfidentNASv3 (CIFAR-10) (Google Brain, 2016-11-05): 2.2e21 FLOP, confidence LikelyNAS with base 8 and shared embeddings (Google Brain, 2016-11-05): 1.1e16 FLOP, confidence ConfidentBIDAF (University of Washington,Allen Institute for AI, 2016-11-05): 3.5e18 FLOP, confidence ConfidentResNeXt-101 (64×4d) (University of California San Diego,Facebook, 2016-11-16): 1.2e19 FLOP, confidence ConfidentPolyNet (Chinese University of Hong Kong (CUHK), 2016-11-17): 6.4e19 FLOP, confidence LikelyHR-ResNet101 (Carnegie Mellon University (CMU), 2016-12-13): 7.1e18 FLOP, confidence ConfidentEnhanceNet (Max Planck Institute for Intelligent Systems, 2016-12-23): 1.3e17 FLOP, confidence ConfidentDeepStack (University of Alberta,Charles University,Czech Technical University, 2017-01-06): 1.4e19 FLOP, confidence SpeculativeMoE-Multi (Jagiellonian University,Google Brain, 2017-01-23): 9.4e19 FLOP, confidence ConfidentTransformer (Google Research,Google Brain, 2017-06-12): 7.4e18 FLOP, confidence ConfidentDeepLoc (Technical University of Denmark,University of Copenhagen, 2017-07-07): 5.8e17 FLOP, confidence LikelyJFT (Google Research,Carnegie Mellon University (CMU), 2017-07-10): 8.4e20 FLOP, confidence ConfidentConvS2S (ensemble of 8 models) (Meta AI, 2017-07-25): 5.6e19 FLOP, confidence LikelyRetinaNet-R101 (Facebook AI Research, 2017-08-07): 2.1e18 FLOP, confidence ConfidentAWD-LSTM - 3-layer LSTM (tied) + continuous cache pointer (WT2) (Salesforce Research, 2017-08-07): 3.0e17 FLOP, confidence ConfidentOpenAI TI7 DOTA 1v1 (OpenAI, 2017-08-11): 6.0e20 FLOP, confidence ConfidentEI-REHN-1000D (Korea Advanced Institute of Science and Technology (KAIST), 2017-08-14): 1.1e16 FLOP, confidence ConfidentLibratus (Carnegie Mellon University (CMU), 2017-08-19): 5.5e20 FLOP, confidence LikelyGL-LWGC-AWD-MoS-LSTM + dynamic evaluation (WT2) (Ben-Gurion University of the Negev, 2017-08-29): 4.6e17 FLOP, confidence ConfidentPyramidNet (Korea Advanced Institute of Science and Technology (KAIST), 2017-09-06): 2.3e15 FLOP, confidence LikelyISS (Duke University,Microsoft, 2017-09-15): 3.4e15 FLOP, confidence ConfidentAWD-LSTM+WT+Cache+IOG (WT2) (NTT Communication Science Laboratories, 2017-09-26): 3.2e15 FLOP, confidence ConfidentAlphaGo Zero (DeepMind, 2017-10-18): 6.5e20 FLOP, confidence ConfidentAlphaGo Master (DeepMind, 2017-10-19): 3.4e20 FLOP, confidence LikelyFraternal dropout + AWD-LSTM 3-layer (WT2) (Jagiellonian University,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),University of Montreal / Université de Montréal, 2017-10-31): 3.1e17 FLOP, confidence LikelyAWD-LSTM-MoS + dynamic evaluation (WT2, 2017) (Carnegie Mellon University (CMU), 2017-11-10): 3.4e18 FLOP, confidence LikelyAlphaZero (DeepMind, 2017-12-05): 1.1e20 FLOP, confidence LikelyQRNN (Salesforce Research, 2018-02-01): 6.9e17 FLOP, confidence LikelyELMo (University of Washington,Allen Institute for AI, 2018-02-01): 3.3e15 FLOP, confidence SpeculativeIMPALA (DeepMind, 2018-02-05): 1.7e20 FLOP, confidence Confident4 layer QRNN (h=2500) (Salesforce Research, 2018-03-22): 5.9e17 FLOP, confidence LikelyYOLOv3 (University of Washington, 2018-04-08): 1.3e19 FLOP, confidence LikelyResNeXt-101 32x48d (Facebook, 2018-05-02): 8.7e21 FLOP, confidence ConfidentDropout-LSTM+Noise(Bernoulli) (WT2) (Columbia University,New York University (NYU),Princeton University, 2018-05-03): 1.3e17 FLOP, confidence LikelyaLSTM(depth-2)+RecurrentPolicy (WT2) (University of Manchester,Alan Turing Institute, 2018-05-22): 7.3e16 FLOP, confidence ConfidentGPT-1 (OpenAI, 2018-06-01): 1.8e19 FLOP, confidence LikelyFTW (For The Win) (DeepMind, 2018-07-03): 3.5e19 FLOP, confidence SpeculativeBig-Little Net (speech) (IBM, 2018-07-10): 4.3e17 FLOP, confidence SpeculativeBig-Little Net (IBM, 2018-07-10): 2.5e17 FLOP, confidence LikelyBig Transformer for Back-Translation (Facebook AI Research,Google Brain, 2018-08-28): 4.8e20 FLOP, confidence Likely(ensemble): AWD-LSTM-DOC (fin) × 5 (WT2) (NTT Communication Science Laboratories,Tohoku University, 2018-08-30): 6.7e17 FLOP, confidence ConfidentTransformer + Simple Recurrent Unit (ASAPP,Cornell University,Google,Princeton University, 2018-09-17): 1.1e19 FLOP, confidence ConfidentLSTM+NeuralCache (KU Leuven,ESAT - PSI,Apple, 2018-09-24): 9.8e14 FLOP, confidence LikelyTransformer (Adaptive Input Embeddings) WT103 (Facebook AI Research, 2018-09-28): 4.5e19 FLOP, confidence ConfidentBERT-Large (Google, 2018-10-11): 2.9e20 FLOP, confidence ConfidentTrellisNet (Carnegie Mellon University (CMU),Bosch Center for Artificial Intelligence,Intel Labs, 2018-10-15): 2.8e18 FLOP, confidence ConfidentMesh-TensorFlow Transformer 4.9B (language) (Google Brain, 2018-11-05): 1.6e20 FLOP, confidence ConfidentMesh-TensorFlow Transformer 2.9B (translation) (Google Brain, 2018-11-05): 6.8e19 FLOP, confidence LikelyFine-tuned-AWD-LSTM-DOC (fin) (Samsung R&D Institute Russia, 2018-11-12): 5.2e16 FLOP, confidence ConfidentMulti-cell LSTM (University of Hyderabad, 2018-11-15): 2.0e15 FLOP, confidence LikelyStyleGAN (Nvidia, 2018-12-12): 3.9e16 FLOP, confidence ConfidentTransformer-XL (257M) (Carnegie Mellon University (CMU),Google Brain, 2019-01-09): 3.8e20 FLOP, confidence SpeculativeHanabi 4 player (DeepMind,University of Oxford,Carnegie Mellon University (CMU),Google Brain, 2019-02-01): 4.3e18 FLOP, confidence ConfidentGPT-2 (1.5B) (OpenAI, 2019-02-14): 1.9e21 FLOP, confidence SpeculativeKataGo (Jane Street, 2019-02-27): 2.3e19 FLOP, confidence SpeculativeSciBERT (Allen Institute for AI, 2019-03-26): 8.9e19 FLOP, confidence ConfidentCross-lingual alignment (Tel Aviv University,Massachusetts Institute of Technology (MIT), 2019-04-04): 2.6e18 FLOP, confidence SpeculativeWeNet (Penn Treebank) (Amazon, 2019-04-08): 7.3e17 FLOP, confidence LikelyBERT-Large-CAS (PTB+WT2+WT103) (Amazon, 2019-04-20): 1.5e20 FLOP, confidence LikelyMuseNet (OpenAI, 2019-04-25): 2.2e20 FLOP, confidence LikelyAWD-LSTM-DRILL + dynamic evaluation† (WT2) (IDIAP, 2019-05-14): 4.1e17 FLOP, confidence LikelyDLRM-2020 (Facebook AI, 2019-05-31): 4.0e18 FLOP, confidence ConfidentXLNet (Carnegie Mellon University (CMU),Google Brain, 2019-06-01): 6.2e21 FLOP, confidence ConfidentTransformer-XL Large + Phrase Induction (Massachusetts Institute of Technology (MIT),University of Illinois Urbana-Champaign (UIUC), 2019-06-04): 3.8e20 FLOP, confidence SpeculativeAWD-LSTM + MoS + Partial Shuffled (University of Texas at Austin, 2019-06-10): 3.1e17 FLOP, confidence ConfidentRoBERTa Large (Facebook,University of Washington, 2019-07-01): 8.5e21 FLOP, confidence ConfidentPluribus (Facebook AI Research, 2019-07-11): 6.6e16 FLOP, confidence LikelytrRosetta (Nankai University,University of Washington,Tianjin University,Harvard University, 2019-08-22): 3.8e19 FLOP, confidence ConfidentUDSMProt (Fraunhofer Heinrich Hertz Institute, 2019-09-04): 6.4e17 FLOP, confidence LikelyMegatron-LM (8.3B) (Nvidia, 2019-09-17): 9.1e21 FLOP, confidence LikelyMegatron-LM (1.2B) (Nvidia, 2019-09-17): 1.1e21 FLOP, confidence ConfidentMegatron-BERT (Nvidia, 2019-09-17): 2.2e22 FLOP, confidence ConfidentDistilBERT (Hugging Face, 2019-10-02): 1.2e19 FLOP, confidence ConfidentAlphaX-1 (Facebook AI Research,Brown University, 2019-10-02): 8.9e17 FLOP, confidence ConfidentT5-11B (Google, 2019-10-23): 3.3e22 FLOP, confidence ConfidentT5-3B (Google, 2019-10-23): 9.0e21 FLOP, confidence ConfidentAlphaStar (DeepMind, 2019-10-30): 1.1e23 FLOP, confidence ConfidentBase LM + kNN LM + Continuous Cache (Stanford University,Facebook AI Research, 2019-11-01): 3.1e19 FLOP, confidence LikelyXLM-RoBERTa (Facebook AI, 2019-11-05): 2.1e22 FLOP, confidence ConfidentSandwich Transformer (Allen Institute for AI,Facebook AI Research, 2019-11-10): 2.4e19 FLOP, confidence SpeculativeCamemBERT (Facebook,INRIA,Sorbonne University, 2019-11-10): 8.3e20 FLOP, confidence ConfidentNoisy Student (L2) (Carnegie Mellon University (CMU),Google, 2019-11-11): 2.6e22 FLOP, confidence ConfidentMuZero (DeepMind, 2019-11-19): 4.8e19 FLOP, confidence SpeculativeTransformer-XL DeFINE (141M) (University of Washington,Allen Institute for AI, 2019-11-27): 1.7e18 FLOP, confidence SpeculativeMMLSTM (WT-2) (Beijing University of Posts and Telecommunications,University of West London, 2019-12-05): 1.9e17 FLOP, confidence LikelyMMLSTM (PTB) (Beijing University of Posts and Telecommunications,University of West London, 2019-12-05): 5.8e16 FLOP, confidence LikelyOpenAI Five Rerun (OpenAI, 2019-12-13): 1.3e22 FLOP, confidence ConfidentOpenAI Five (OpenAI, 2019-12-13): 6.7e22 FLOP, confidence ConfidentDD-PPO (Georgia Institute of Technology,Facebook AI Research,Oregon State University,Simon Fraser University, 2019-12-19): 7.8e20 FLOP, confidence LikelyAlphaFold (DeepMind, 2020-01-15): 1.0e20 FLOP, confidence SpeculativeContextNet + Noisy Student (Google, 2020-01-19): 8.2e21 FLOP, confidence ConfidentMeena (Google Brain, 2020-01-28): 1.1e23 FLOP, confidence ConfidentTaLK Convolution (Carleton University, 2020-02-08): 2.7e19 FLOP, confidence LikelyALBERT-xxlarge (Toyota Technological Institute at Chicago,Google, 2020-02-09): 2.4e21 FLOP, confidence SpeculativeTuring-NLG (Microsoft, 2020-02-13): 1.6e22 FLOP, confidence LikelyFFN SwiGLU (Google, 2020-02-14): 3.4e19 FLOP, confidence ConfidentFeedback Transformer (LORIA,University of Lorraine,Facebook AI Research, 2020-02-21): 7.7e18 FLOP, confidence SpeculativeTransformerXL + spectrum control (University of California Los Angeles (UCLA),JD.com, 2020-03-11): 2.6e19 FLOP, confidence SpeculativeTensor-Transformer(1core)+PN (WT103) (University of California (UC) Berkeley, 2020-03-17): 1.6e18 FLOP, confidence ConfidentELECTRA (Stanford University,Google,Google Brain, 2020-03-23): 3.1e21 FLOP, confidence LikelyMetNet (Google, 2020-03-24): 9.5e18 FLOP, confidence SpeculativeOnce for All (MIT-IBM Watson AI Lab,Massachusetts Institute of Technology (MIT),IBM, 2020-04-29): 6.2e20 FLOP, confidence ConfidentUnifiedQA (Allen Institute for AI,University of Washington, 2020-05-02): 1.6e19 FLOP, confidence ConfidentDETR (Facebook, 2020-05-26): 4.0e20 FLOP, confidence ConfidentGPT-3 175B (davinci) (OpenAI, 2020-05-28): 3.1e23 FLOP, confidence ConfidentGShard (dense) (Google, 2020-06-30): 4.8e22 FLOP, confidence ConfidentDeLighT (University of Washington,Allen Institute for AI,Facebook AI Research, 2020-08-03): 3.8e18 FLOP, confidence LikelyERNIE-GEN (large) (Baidu, 2020-08-06): 2.0e20 FLOP, confidence SpeculativeProBERTa (University of Illinois Urbana-Champaign (UIUC),Reed College, 2020-09-01): 9.7e18 FLOP, confidence ConfidentLUKE (University of Washington,National Institute of Informatics, 2020-10-02): 1.8e22 FLOP, confidence LikelymT5-XXL (Google,Google Research, 2020-10-20): 8.2e22 FLOP, confidence ConfidentConformer + Wav2vec 2.0 + Noisy Student (Google,Google Research,Google Brain, 2020-10-20): 7.6e21 FLOP, confidence ConfidentGerman ELECTRA Large (deepset,Bayerische Staatsbibliothek Muenchen, 2020-10-21): 1.4e21 FLOP, confidence Confidentwave2vec 2.0 LARGE (Facebook, 2020-10-22): 3.9e21 FLOP, confidence ConfidentViT-Huge/14 (Google Brain,Google Research, 2020-10-22): 4.3e21 FLOP, confidence ConfidentKEPLER (Tsinghua University,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),HEC,CIFAR AI Research,Princeton University,University of Montreal / Université de Montréal, 2020-11-23): 1.7e21 FLOP, confidence ConfidentAlphaFold 2 (DeepMind, 2020-11-30): 3.0e21 FLOP, confidence LikelyCPM-Large (Tsinghua University,Beijing Academy of Artificial Intelligence / BAAI, 2020-12-01): 2.6e20 FLOP, confidence ConfidentESM1b (Facebook AI Research,New York University (NYU), 2020-12-15): 5.1e21 FLOP, confidence ConfidentDensePhrases (Korea University,Princeton University, 2020-12-23): 2.1e18 FLOP, confidence SpeculativeCT-MoS (WT2) (Google,National Tsing Hua University, 2020-12-25): 5.4e17 FLOP, confidence ConfidentERNIE-Doc (247M) (Baidu, 2020-12-31): 3.0e19 FLOP, confidence ConfidentDALL-E (OpenAI, 2021-01-05): 4.7e22 FLOP, confidence LikelyCLIP (ViT L/14@336px) (OpenAI, 2021-01-05): 1.0e22 FLOP, confidence ConfidentSwitch (Google, 2021-01-11): 8.2e22 FLOP, confidence ConfidentDeiT-B (Meta AI,Sorbonne University, 2021-01-15): 7.9e19 FLOP, confidence ConfidentDLWP (University of Washington,Microsoft Research, 2021-02-09): 5.7e18 FLOP, confidence ConfidentMSA Transformer (Facebook AI Research,University of California (UC) Berkeley,New York University (NYU), 2021-02-13): 5.5e21 FLOP, confidence LikelySRU++ Large (ASAPP, 2021-02-24): 2.1e19 FLOP, confidence ConfidentMeta Pseudo Labels (Google Brain,Google AI, 2021-03-01): 4.8e22 FLOP, confidence LikelyM6-T (Alibaba, 2021-03-05): 5.5e21 FLOP, confidence LikelyGenerative BST (Facebook AI Research, 2021-03-05): 1.4e22 FLOP, confidence ConfidentPLUG (Alibaba, 2021-04-19): 3.6e22 FLOP, confidence ConfidentProtT5-XL-U50 (Technical University of Munich,Med AI Technology,Nvidia,Oak Ridge National Laboratory,Google,Seoul National University, 2021-05-04): 1.9e22 FLOP, confidence ConfidentProtBERT-BFD (Technical University of Munich,Nvidia,Seoul National University,Google,Oak Ridge National Laboratory,Med AI Technology, 2021-05-04): 3.9e22 FLOP, confidence ConfidentADM (OpenAI, 2021-05-11): 6.2e21 FLOP, confidence ConfidentMedBERT (Peng Cheng Laboratory,University of Texas at Houston, 2021-05-20): 9.5e18 FLOP, confidence LikelyTransformer local-attention (NesT-B) (Google Cloud,Google Research, 2021-05-26): 2.4e19 FLOP, confidence ConfidentCogView (Tsinghua University,Alibaba DAMO Academy, 2021-05-26): 2.7e22 FLOP, confidence LikelyByT5-XXL (Google,Google Research, 2021-05-28): 8.1e22 FLOP, confidence LikelyViT-G/14 (Google Brain,Google Research, 2021-06-08): 5.8e22 FLOP, confidence ConfidentEMDR (Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),McGill University,DeepMind, 2021-06-09): 1.9e21 FLOP, confidence ConfidentCoAtNet (Google,Google Research,Google Brain, 2021-06-09): 4.3e22 FLOP, confidence ConfidentDeBERTa (Microsoft, 2021-06-10): 2.6e22 FLOP, confidence ConfidentDenoising Diffusion Probabilistic Models (LSUN Bedroom) (University of California (UC) Berkeley, 2021-06-11): 7.8e19 FLOP, confidence ConfidentALIGN (Google Research, 2021-06-11): 2.6e22 FLOP, confidence ConfidentStyleGAN3-T (Nvidia,Aalto University, 2021-06-21): 1.7e21 FLOP, confidence ConfidentStyleGAN3-R (Nvidia,Aalto University, 2021-06-21): 2.4e21 FLOP, confidence ConfidentEfficientNetV2-XL (Google,Google Brain, 2021-06-23): 9.6e19 FLOP, confidence ConfidentFold2Seq (IBM,Texas A&M, 2021-06-24): 1.4e17 FLOP, confidence SpeculativeAdaptive Input Transformer + RD (Microsoft Research Asia,Soochow University, 2021-06-28): 8.6e19 FLOP, confidence LikelyERNIE 3.0 (Baidu, 2021-07-05): 2.3e22 FLOP, confidence ConfidentCodex (OpenAI, 2021-07-07): 7.3e22 FLOP, confidence LikelyHuBERT (Facebook AI Research, 2021-07-27): 5.5e21 FLOP, confidence SpeculativeGOAT (DeepMind, 2021-07-27): 2.4e22 FLOP, confidence SpeculativeSEER (Facebook AI Research,INRIA, 2021-07-29): 1.8e22 FLOP, confidence ConfidentYOLOX-X (Megvii Inc, 2021-08-06): 6.3e20 FLOP, confidence LikelyZidong Taichu (Chinese Academy of Sciences,Wuhan AI Computing Center, 2021-08-11): 8.0e20 FLOP, confidence ConfidentJurassic-1-Jumbo (AI21 Labs, 2021-08-11): 3.7e23 FLOP, confidence ConfidentDNABERT (Northeastern University, 2021-08-15): 1.1e20 FLOP, confidence ConfidentXLMR-XXL (Facebook AI Research, 2021-08-17): 3.4e22 FLOP, confidence ConfidentFLAN 137B (Google Research, 2021-09-03): 2.0e24 FLOP, confidence ConfidentPermuteFormer (Peking University, 2021-09-06): 2.8e18 FLOP, confidence SpeculativeHyperCLOVA 204B (NAVER, 2021-09-10): 2.0e23 FLOP, confidence SpeculativePLATO-XL (Baidu, 2021-09-20): 9.9e21 FLOP, confidence ConfidentTuring ULRv5 (Microsoft, 2021-09-28): 2.9e22 FLOP, confidence ConfidentAlphaFold-Multimer (Google DeepMind,DeepMind, 2021-10-04): 4.3e21 FLOP, confidence ConfidentMegatron-Turing NLG 530B (Microsoft,Nvidia, 2021-10-11): 8.6e23 FLOP, confidence ConfidentYuan 1.0 (Inspur, 2021-10-12): 3.5e23 FLOP, confidence Confidentbase LM+GNN+kNN (Shannon.AI,Nanjing University,Nanyang Technological University,Zhejiang University (ZJU), 2021-10-17): 5.3e19 FLOP, confidence ConfidentS4 (Stanford University, 2021-10-31): 7.8e19 FLOP, confidence LikelyProjected GAN (Heidelberg University, 2021-11-01): 1.1e19 FLOP, confidence ConfidentCodeT5-base (Salesforce,Nanyang Technological University, 2021-11-01): 1.6e21 FLOP, confidence LikelyMasked Autoencoders ViT-H (Facebook AI Research, 2021-11-11): 4.6e20 FLOP, confidence SpeculativeSwin Transformer V2 (SwinV2-G) (Microsoft Research Asia, 2021-11-18): 1.1e21 FLOP, confidence ConfidentBASIC-L (Google, 2021-11-19): 4.1e22 FLOP, confidence LikelyFlorence (Microsoft, 2021-11-22): 4.8e22 FLOP, confidence ConfidentNÜWA (Microsoft Research,Peking University, 2021-11-24): 7.2e21 FLOP, confidence ConfidentStudent of Games (DeepMind, 2021-12-06): 3.7e22 FLOP, confidence SpeculativeGopher (280B) (DeepMind, 2021-12-08): 6.3e23 FLOP, confidence ConfidentGLaM (Google, 2021-12-13): 3.6e23 FLOP, confidence ConfidentContriever (Meta AI,University College London (UCL),PSL University,Université Grenoble Alpes, 2021-12-16): 1.6e20 FLOP, confidence LikelyXGLM-7.5B (Meta AI,Facebook AI Research, 2021-12-20): 2.3e22 FLOP, confidence ConfidentERNIE 3.0 Titan (Baidu,Peng Cheng Laboratory, 2021-12-23): 1.0e24 FLOP, confidence ConfidentDetic (Meta AI,University of Texas at Austin, 2022-01-07): 2.3e19 FLOP, confidence SpeculativeInstructGPT 175B (OpenAI, 2022-01-27): 3.2e23 FLOP, confidence ConfidentAlphaCode (DeepMind, 2022-02-02): 2.4e23 FLOP, confidence ConfidentRETRO-7B (DeepMind, 2022-02-07): 1.7e22 FLOP, confidence ConfidentGPT-NeoX-20B (EleutherAI, 2022-02-09): 9.3e22 FLOP, confidence ConfidentProteinBERT (Hebrew University of Jerusalem,Ben-Gurion University of the Negev,Deep Trading, 2022-02-10): 6.5e19 FLOP, confidence ConfidentLaMDA (Google, 2022-02-10): 3.6e23 FLOP, confidence ConfidentST-MoE (Google,Google Brain,Google Research, 2022-02-17): 2.9e23 FLOP, confidence LikelyFourCastNet (Nvidia,NERSC, Lawrence Berkeley National Laboratory,University of Michigan,Rice University,California Institute of Technology,Purdue University, 2022-02-22): 3.5e20 FLOP, confidence ConfidentPolyCoder (Carnegie Mellon University (CMU), 2022-02-26): 1.1e21 FLOP, confidence LikelyStatement Curriculum Learning (OpenAI, 2022-03-02): 1.8e22 FLOP, confidence LikelyViT-G (model soup) (University of Washington,Columbia University,Google,Meta AI,Tel Aviv University, 2022-03-10): 3.4e21 FLOP, confidence ConfidentGPT-3.5 (davinci-002) (OpenAI, 2022-03-15): 2.6e24 FLOP, confidence SpeculativeSegatron-XL large, M=384 + HCP (Microsoft Research,University of Waterloo, 2022-03-21): 2.6e19 FLOP, confidence ConfidentMake-A-Scene (Meta AI, 2022-03-24): 6.4e21 FLOP, confidence LikelyChinchilla (DeepMind, 2022-03-29): 5.8e23 FLOP, confidence ConfidentPaLM (540B) (Google Research, 2022-04-04): 2.5e24 FLOP, confidence ConfidentDALL·E 2 (OpenAI, 2022-04-06): 3.4e23 FLOP, confidence SpeculativeBERT-RBP (Waseda University, 2022-04-07): 1.4e20 FLOP, confidence ConfidentStable Diffusion (LDM-KL-8-G) (Runway,Ludwig Maximilian University of Munich,Heidelberg University, 2022-04-13): 5.0e22 FLOP, confidence ConfidentSparse all-MLP (Meta AI, 2022-04-14): 5.3e20 FLOP, confidence ConfidentFlamingo (DeepMind, 2022-04-29): 2.2e23 FLOP, confidence ConfidentOPT-175B (Meta AI, 2022-05-02): 4.3e23 FLOP, confidence ConfidentUL2 (Google Research,Google Brain, 2022-05-10): 1.2e23 FLOP, confidence ConfidentGato (DeepMind, 2022-05-12): 4.0e21 FLOP, confidence ConfidentImagen (Google Brain, 2022-05-23): 1.5e22 FLOP, confidence LikelyTranception (University of Oxford,Harvard Medical School,Cohere, 2022-05-27): 7.2e21 FLOP, confidence ConfidentGPT-2 Medium (FlashAttention) (Stanford University,University at Buffalo, 2022-05-27): 8.9e20 FLOP, confidence ConfidentDITTO (Tsinghua University,Apple,Westlake University,Chinese University of Hong Kong (CUHK), 2022-06-06): 3.3e18 FLOP, confidence ConfidentCoCa (Google Research, 2022-06-14): 7.3e22 FLOP, confidence ConfidentParti (Google Research, 2022-06-22): 5.1e23 FLOP, confidence ConfidentProGen2-xlarge (Salesforce Research,Columbia University,Johns Hopkins University, 2022-06-27): 1.4e22 FLOP, confidence ConfidentMinerva (540B) (Google, 2022-06-29): 2.7e24 FLOP, confidence ConfidentCodeT5-large (Salesforce, 2022-07-05): 2.7e21 FLOP, confidence LikelyNLLB (Meta AI, 2022-07-06): 1.8e22 FLOP, confidence ConfidentBLOOM-176B (Hugging Face,BigScience, 2022-07-11): 3.7e23 FLOP, confidence ConfidentESM2-15B (Meta AI,New York University (NYU),Stanford University,Massachusetts Institute of Technology (MIT), 2022-07-21): 7.4e22 FLOP, confidence ConfidentOmegaPLM (Massachusetts Institute of Technology (MIT),Westlake University, 2022-07-22): 1.0e22 FLOP, confidence ConfidentAlexaTM 20B (Amazon, 2022-08-02): 2.0e23 FLOP, confidence ConfidentGLM-130B (Tsinghua University, 2022-08-04): 3.5e23 FLOP, confidence ConfidentBlenderBot 3 (McGill University,Meta AI,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms), 2022-08-10): 4.3e23 FLOP, confidence LikelyBEIT-3 (Microsoft, 2022-08-22): 7.0e19 FLOP, confidence LikelyPaLI (Google, 2022-09-14): 1.7e23 FLOP, confidence LikelyWhisper (OpenAI, 2022-09-21): 4.2e21 FLOP, confidence LikelyDiffDock (Massachusetts Institute of Technology (MIT), 2022-10-04): 7.2e19 FLOP, confidence LikelyAlphaTensor (DeepMind, 2022-10-05): 7.1e20 FLOP, confidence ConfidentGenSLM (University of Chicago,Nvidia,Harvard University,Cerebras Systems,Technical University of Munich,California Institute of Technology, 2022-10-11): 1.4e21 FLOP, confidence ConfidentU-PaLM (540B) (Google, 2022-10-20): 2.5e24 FLOP, confidence ConfidentFlan-PaLM 540B (Google, 2022-10-20): 2.5e24 FLOP, confidence ConfidenteDiff-I (Nvidia, 2022-11-02): 5.5e19 FLOP, confidence LikelyMogrifier RLSTM (WT2) (DeepMind, 2022-11-03): 1.4e17 FLOP, confidence ConfidentInternImage (Shanghai AI Lab,Tsinghua University,Nanjing University,SenseTime,Chinese University of Hong Kong (CUHK), 2022-11-10): 2.4e21 FLOP, confidence ConfidentEVA-01 (Beijing Academy of Artificial Intelligence / BAAI,Huazhong University of Science and Technology,Zhejiang University (ZJU),Beijing Institute of Technology, 2022-11-14): 1.5e22 FLOP, confidence ConfidentGalactica (Meta AI, 2022-11-16): 3.2e23 FLOP, confidence LikelyFusion in Encoder (Samsung, 2022-11-18): 1.3e20 FLOP, confidence LikelyAR-LDM (Alibaba,University of Waterloo,Vector Institute, 2022-11-20): 5.1e20 FLOP, confidence ConfidentDiscriminator Guidance (Korea Advanced Institute of Science and Technology (KAIST),NAVER, 2022-11-28): 2.2e20 FLOP, confidence ConfidentVega v2 (Wuhan University,JD Explore Academy,Shanghai AI Lab,Nanyang Technological University,Washington University in St Louis,Chongqing University of Posts and Telecommunications,University of Sydney, 2022-12-04): 7.8e22 FLOP, confidence ConfidentCaLM (University of Oxford, 2022-12-19): 2.9e19 FLOP, confidence LikelyHybrid H3-2.7B (Stanford University,University at Buffalo, 2022-12-28): 6.5e21 FLOP, confidence LikelyVALL-E (Microsoft, 2023-01-05): 1.0e19 FLOP, confidence SpeculativeDreamerV3 (DeepMind,University of Toronto, 2023-01-10): 2.2e20 FLOP, confidence LikelyNucleotide Transformer (Nvidia,Technical University of Munich,InstaDeep, 2023-01-15): 8.1e21 FLOP, confidence LikelyAnkh_large (Technical University of Munich,Columbia University, 2023-01-16): 6.5e21 FLOP, confidence ConfidentDDPM-IP (CelebA) (Utrecht University, 2023-01-27): 3.5e20 FLOP, confidence LikelyBLIP-2 (Q-Former) (Salesforce Research, 2023-01-30): 1.2e21 FLOP, confidence ConfidentViT-22B (Google, 2023-02-10): 1.9e23 FLOP, confidence ConfidentLLaMA-65B (Meta AI, 2023-02-24): 5.5e23 FLOP, confidence ConfidentDiT-XL/2 (New York University (NYU),University of California (UC) Berkeley, 2023-03-02): 6.0e20 FLOP, confidence ConfidentAudioGen (Meta AI,Hebrew University of Jerusalem, 2023-03-05): 9.5e21 FLOP, confidence LikelyGPT-4 (Mar 2023) (OpenAI, 2023-03-15): 2.1e25 FLOP, confidence LikelyFalcon-40B (Technology Innovation Institute, 2023-03-15): 2.4e23 FLOP, confidence ConfidentPanGu-Σ (Huawei Noah's Ark Lab, 2023-03-20): 4.7e23 FLOP, confidence ConfidentSigLIP 400M (Google DeepMind, 2023-03-27): 4.9e21 FLOP, confidence ConfidentVideoMAE V2 (Nanjing University,Shenzhen Institute of Advanced Technology,Shanghai AI Lab, 2023-03-29): 9.7e21 FLOP, confidence ConfidentBloombergGPT (Bloomberg,Johns Hopkins University, 2023-03-30): 2.4e23 FLOP, confidence ConfidentSegment Anything Model (Meta AI, 2023-04-05): 7.8e21 FLOP, confidence ConfidentIncoder-6.7B (Facebook AI Research,University of Washington,University of California (UC) Berkeley,Carnegie Mellon University (CMU),Toyota Technological Institute at Chicago, 2023-04-09): 3.0e21 FLOP, confidence ConfidentDINOv2 (Facebook AI Research,INRIA, 2023-04-14): 7.4e21 FLOP, confidence ConfidentLLaVA (University of Wisconsin Madison,Microsoft Research,Columbia University, 2023-04-17): 7.8e22 FLOP, confidence ConfidentStarCoder (Hugging Face,ServiceNow,Northeastern University,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),Carnegie Mellon University (CMU),Johns Hopkins University,Leipzig University,ScaDS.AI,Queen Mary University of London,Roblox,Sea AI Lab,Technion - Israel Institute of Technology,Monash University,CSIRO,Data61,McGill University,Saama,University of British Columbia (UBC),Massachusetts Institute of Technology (MIT),Technical University of Munich,IBM,University of Vermont,UnfoldML,SAP,University of Notre Dame,Columbia University,New York University (NYU),University of Allahabad,Discover Dollar,Toloka,Telefonica,Stanford University,Weizmann Institute of Science,Alan Turing Institute,Wellesley College,EleutherAI,Forschungszentrum Julich, 2023-05-09): 8.5e22 FLOP, confidence ConfidentPaLM 2 (Google, 2023-05-10): 7.3e24 FLOP, confidence LikelyInstructBLIP (Salesforce Research,Hong Kong University of Science and Technology (HKUST),Nanyang Technological University, 2023-05-11): 1.9e20 FLOP, confidence ConfidentONE-PEACE (Alibaba,Huazhong University of Science and Technology, 2023-05-18): 1.8e20 FLOP, confidence SpeculativePaLI-X (Google Research, 2023-05-29): 5.6e23 FLOP, confidence LikelyGPT-4 (Jun 2023) (OpenAI, 2023-06-13): 2.1e25 FLOP, confidence LikelyHyenaDNA (Stanford University,Harvard University,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),University of Montreal / Université de Montréal, 2023-06-27): 1.8e21 FLOP, confidence ConfidentPangu-Weather (Huawei, 2023-07-05): 4.0e22 FLOP, confidence ConfidentxTrimoPGLM -100B (Tsinghua University,BioMap Research, 2023-07-06): 6.2e23 FLOP, confidence ConfidentInternLM (Shanghai AI Lab,SenseTime, 2023-07-06): 10.0e23 FLOP, confidence ConfidentClaude 2 (Anthropic, 2023-07-11): 3.9e24 FLOP, confidence SpeculativeLlama 2-70B (Meta AI, 2023-07-18): 8.1e23 FLOP, confidence ConfidentLlama 2-7B (Meta AI, 2023-07-18): 8.4e22 FLOP, confidence ConfidentAudioLM (Google Research, 2023-07-26): 3.9e18 FLOP, confidence SpeculativeGGNN (Westlake University,Tsinghua University,Toyota Technological Institute at Chicago, 2023-08-05): 7.6e21 FLOP, confidence ConfidentPeptideBERT (Carnegie Mellon University (CMU), 2023-08-28): 4.9e16 FLOP, confidence ConfidentJais (Cerebras Systems,Mohamed bin Zayed University of Artificial Intelligence (MBZUAI),Inception G42, 2023-08-29): 4.9e22 FLOP, confidence ConfidentSwift (Intel Labs, 2023-08-30): 5.3e16 FLOP, confidence LikelyFalcon-180B (Technology Innovation Institute, 2023-09-06): 3.8e24 FLOP, confidence ConfidentAmazon Titan (Amazon, 2023-09-28): 4.8e24 FLOP, confidence LikelyFinGPT-13B (University of California Los Angeles (UCLA),Columbia University,New York University (NYU), 2023-10-07): 1.6e23 FLOP, confidence LikelyRoseTTAFold All-Atom (RFAA) (University of Washington,Seoul National University,University of Sheffield, 2023-10-09): 2.1e20 FLOP, confidence ConfidentCODEFUSION (Python) (Microsoft,Microsoft Research, 2023-10-26): 7.9e18 FLOP, confidence ConfidentChatGLM3-6B (Z.ai (Zhipu AI), 2023-10-27): 5.0e22 FLOP, confidence SpeculativeSkywork-13B (Kunlun Inc., 2023-10-30): 2.5e23 FLOP, confidence ConfidentYi-34B (01.AI, 2023-11-02): 6.1e23 FLOP, confidence ConfidentGrok-1 (xAI, 2023-11-04): 2.9e24 FLOP, confidence LikelyLLaVA 1.5 (University of Wisconsin Madison,Microsoft Research, 2023-11-05): 7.8e22 FLOP, confidence ConfidentCogVLM-17B (Tsinghua University,Z.ai (Zhipu AI),Beihang University, 2023-11-06): 6.3e22 FLOP, confidence ConfidentRoFormer (Zhuiyi Technology, 2023-11-08): 2.2e18 FLOP, confidence ConfidentMultiBand Diffusion (Meta AI,Hebrew University of Jerusalem,LORIA, 2023-11-08): 2.6e19 FLOP, confidence ConfidentVolcano 13B (Korea University,Korea Advanced Institute of Science and Technology (KAIST),LG, 2023-11-13): 4.6e22 FLOP, confidence LikelySPHINX (Llama 2 13B) (Shanghai AI Lab,Chinese University of Hong Kong (CUHK),ShanghaiTech University, 2023-11-13): 3.0e22 FLOP, confidence LikelyGraphCast (Google DeepMind, 2023-11-14): 2.1e22 FLOP, confidence SpeculativeNemotron-3-8B (Nvidia, 2023-11-15): 1.8e23 FLOP, confidence ConfidentInflection-2 (Inflection AI, 2023-11-22): 1.0e25 FLOP, confidence ConfidentQwen-72B (Alibaba, 2023-11-30): 1.3e24 FLOP, confidence ConfidentGemini 1.0 Ultra (Google DeepMind, 2023-12-06): 5.0e25 FLOP, confidence SpeculativeLlama Guard (Meta AI, 2023-12-07): 1.6e23 FLOP, confidence ConfidentMixtral 8x7B (Mistral AI, 2023-12-11): 7.7e23 FLOP, confidence SpeculativeVILA-13B (Nvidia,Massachusetts Institute of Technology (MIT), 2023-12-12): 2.3e21 FLOP, confidence ConfidentFunSearch (Google DeepMind, 2023-12-14): 3.9e23 FLOP, confidence SpeculativeCogAgent (Tsinghua University,Z.ai (Zhipu AI), 2023-12-14): 6.7e22 FLOP, confidence Likelynekomata-14b (rinna, 2023-12-21): 2.6e23 FLOP, confidence ConfidentGQA-8-XXL (Google Research, 2023-12-23): 3.5e22 FLOP, confidence ConfidentQwen1.5-72B (Alibaba, 2024-02-04): 1.3e24 FLOP, confidence ConfidentStable Diffusion 3 (Stability AI, 2024-02-22): 5.0e22 FLOP, confidence ConfidentMegaScale (Production) (ByteDance,Peking University, 2024-02-23): 3.9e24 FLOP, confidence SpeculativeAramco Metabrain AI (Saudi Aramco, 2024-03-04): 1.0e25 FLOP, confidence LikelyInflection-2.5 (Inflection AI, 2024-03-07): 8.0e24 FLOP, confidence LikelyMM1-30B (Apple, 2024-03-14): 4.9e23 FLOP, confidence LikelyDBRX (Databricks, 2024-03-27): 2.6e24 FLOP, confidence ConfidentReka Core (Reka AI, 2024-04-15): 8.4e24 FLOP, confidence SpeculativeLlama 3-70B (Meta AI, 2024-04-18): 7.9e24 FLOP, confidence ConfidentGenCast (Google DeepMind, 2024-05-01): 8.2e20 FLOP, confidence ConfidentVILA1.5-13B (Nvidia,Massachusetts Institute of Technology (MIT), 2024-05-03): 2.3e21 FLOP, confidence ConfidentAlphaFold 3 (Google DeepMind,Isomorphic Labs, 2024-05-08): 4.1e22 FLOP, confidence ConfidentYi-Large (01.AI, 2024-05-13): 1.8e24 FLOP, confidence SpeculativeOcto-Base (University of California (UC) Berkeley,Stanford University,Carnegie Mellon University (CMU),DeepMind, 2024-05-20): 5.8e20 FLOP, confidence ConfidentALLaM adapted 70B (Saudi Data and Artificial Intelligence Authority, 2024-05-21): 1.1e24 FLOP, confidence ConfidentQwen2-72B (Alibaba, 2024-06-07): 3.0e24 FLOP, confidence ConfidentOpenVLA (Stanford University,University of California (UC) Berkeley,Toyota Research Institute,Google DeepMind,Massachusetts Institute of Technology (MIT),Physical Intelligence, 2024-06-13): 1.1e23 FLOP, confidence ConfidentNemotron-4 340B (Nvidia, 2024-06-14): 1.8e25 FLOP, confidence ConfidentDeepSeek-Coder-V2 236B (DeepSeek, 2024-06-17): 1.3e24 FLOP, confidence ConfidentClaude 3.5 Sonnet (Anthropic, 2024-06-20): 2.7e25 FLOP, confidence SpeculativeESM3 (98B) (EvolutionaryScale,University of California (UC) Berkeley, 2024-06-25): 1.1e24 FLOP, confidence ConfidentLlama 3.1-405B (Meta AI, 2024-07-23): 3.8e25 FLOP, confidence ConfidentAFM-server (Apple, 2024-07-29): 4.3e24 FLOP, confidence LikelyAFM-on-device (Apple, 2024-07-29): 4.5e23 FLOP, confidence ConfidentLLaVA-OV-72B (ByteDance,Nanyang Technological University,Chinese University of Hong Kong (CUHK),Hong Kong University of Science and Technology (HKUST), 2024-08-06): 3.0e24 FLOP, confidence ConfidentGrok-2 (xAI, 2024-08-13): 3.0e25 FLOP, confidence ConfidentDeepSeek-V2.5 (DeepSeek, 2024-09-06): 1.8e24 FLOP, confidence ConfidentQwen2.5-32B (Alibaba, 2024-09-17): 3.5e24 FLOP, confidence ConfidentQwen2.5-72B (Alibaba, 2024-09-19): 7.8e24 FLOP, confidence ConfidentQwen2.5 Instruct (72B) (Alibaba, 2024-09-19): 7.9e24 FLOP, confidence ConfidentTelechat2-115B (China Telecom, 2024-09-20): 6.9e24 FLOP, confidence ConfidentLlama 3.2 11B (Meta AI, 2024-09-24): 5.8e23 FLOP, confidence ConfidentMovie Gen Video (Meta AI, 2024-10-04): 1.6e24 FLOP, confidence ConfidentRDT-1B (Tsinghua University, 2024-10-10): 4.1e22 FLOP, confidence LikelyLlama-3.1-Nemotron-70B-Instruct (Nvidia,Meta AI, 2024-10-12): 7.9e24 FLOP, confidence ConfidentCHAI-1 (Chai discovery, 2024-10-15): 7.8e21 FLOP, confidence ConfidentYi-Lightning (01.AI, 2024-10-16): 1.5e24 FLOP, confidence ConfidentNVLM-X 72B (Nvidia, 2024-10-22): 3.0e24 FLOP, confidence LikelyNVLM-H 72B (Nvidia, 2024-10-22): 3.0e24 FLOP, confidence LikelyNVLM-D 72B (Nvidia, 2024-10-22): 3.0e24 FLOP, confidence ConfidentDoubao-pro (ByteDance, 2024-10-28): 2.5e25 FLOP, confidence LikelyHunyuan-Large (Tencent, 2024-11-06): 3.5e24 FLOP, confidence ConfidentAmazon Nova Pro (Amazon, 2024-12-03): 6.0e24 FLOP, confidence SpeculativeLlama 3.3 70B (Meta AI, 2024-12-06): 6.9e24 FLOP, confidence ConfidentEXAONE 3.5 32B (LG AI Research, 2024-12-09): 1.3e24 FLOP, confidence ConfidentDeepSeek-V3 (DeepSeek, 2024-12-24): 3.3e24 FLOP, confidence ConfidentEurus-2-7B-PRIME (Tsinghua University,University of Illinois Urbana-Champaign (UIUC),Shanghai AI Lab,Peking University,Shanghai Jiao Tong University,CUHK Shenzhen Research Institute, 2025-01-02): 4.0e20 FLOP, confidence SpeculativeEagle 2 (Nvidia,Nanjing University,Tsinghua University,Hong Kong Polytechnic University,Johns Hopkins University,New York University (NYU), 2025-01-20): 4.7e22 FLOP, confidence ConfidentDeepSeek-R1 (DeepSeek, 2025-01-20): 3.5e24 FLOP, confidence ConfidentGrok 3 (xAI, 2025-02-17): 3.5e26 FLOP, confidence LikelyClaude 3.7 Sonnet (Anthropic, 2025-02-24): 3.4e25 FLOP, confidence LikelyGPT-4.5 (OpenAI, 2025-02-27): 3.8e26 FLOP, confidence LikelyQwQ-32B (Alibaba, 2025-03-06): 3.5e24 FLOP, confidence SpeculativeHunyuan-TurboS (Tencent, 2025-03-11): 5.4e24 FLOP, confidence ConfidentEXAONE Deep 32B (LG AI Research, 2025-03-16): 1.3e24 FLOP, confidence ConfidentDeepSeek-V3 (Mar 2025) (DeepSeek, 2025-03-24): 3.3e24 FLOP, confidence ConfidentLlama 4 Scout (Meta AI, 2025-04-05): 4.1e24 FLOP, confidence LikelyLlama 4 Maverick (Meta AI, 2025-04-05): 2.2e24 FLOP, confidence LikelyLlama 4 Behemoth (preview) (Meta AI, 2025-04-05): 5.2e25 FLOP, confidence LikelyPangu Ultra (Huawei, 2025-04-10): 1.1e25 FLOP, confidence ConfidentQwen3-235B-A22B (Alibaba, 2025-04-29): 4.8e24 FLOP, confidence LikelySeed1.5-VL (ByteDance, 2025-05-11): 1.4e24 FLOP, confidence ConfidentDeepSeek-R1 (May 2025) (DeepSeek, 2025-05-28): 4.0e24 FLOP, confidence ConfidentFGN (Google DeepMind, 2025-06-12): 9.6e21 FLOP, confidence ConfidentGrok 4 (xAI, 2025-07-09): 5.0e26 FLOP, confidence SpeculativeKimi K2 (Moonshot, 2025-07-11): 3.0e24 FLOP, confidence ConfidentEXAONE 4.0 (32B) (LG AI Research, 2025-07-15): 2.7e24 FLOP, confidence ConfidentQwen3-Coder-480B-A35B (Alibaba, 2025-07-22): 1.6e24 FLOP, confidence ConfidentQwen3-235B-A22B-Thinking (Jul 2025) (Alibaba, 2025-07-25): 4.8e24 FLOP, confidence LikelyQwen3-235B-A22B (Jul 2025) (Alibaba, 2025-07-25): 4.8e24 FLOP, confidence LikelyGLM-4.5 (Z.ai (Zhipu AI),Tsinghua University, 2025-07-28): 4.4e24 FLOP, confidence Confidentgpt-oss-120b (OpenAI, 2025-08-05): 4.9e24 FLOP, confidence Confidentgpt-oss-20b (OpenAI, 2025-08-05): 5.5e23 FLOP, confidence ConfidentGPT-5 (OpenAI, 2025-08-07): 6.6e25 FLOP, confidence SpeculativeLongCat-Flash (Meituan Inc, 2025-09-01): 3.7e24 FLOP, confidence LikelyQwen3-Max (Alibaba, 2025-09-05): 1.5e25 FLOP, confidence SpeculativeAgentFounder-30B (Alibaba, 2025-09-16): 6.5e23 FLOP, confidence ConfidentQwen3-Omni-30B-A3B (Alibaba, 2025-09-22): 3.6e22 FLOP, confidence ConfidentGLM-4.6 (Z.ai (Zhipu AI),Tsinghua University, 2025-09-30): 4.4e24 FLOP, confidence LikelyLing-1T (Ant Group, 2025-10-10): 6.0e24 FLOP, confidence ConfidentKimi K2 Thinking (Moonshot, 2025-11-06): 4.2e24 FLOP, confidence LikelyOlmo 3 (Allen Institute for AI, 2025-11-20): 1.1e24 FLOP, confidence ConfidentNemotron 3-Nano-30B-A3B (Nvidia, 2025-12-15): 4.8e23 FLOP, confidence LikelyGLM-4.7 (Z.ai (Zhipu AI), 2025-12-22): 4.4e24 FLOP, confidence LikelySolar Open 100B (Upstage, 2025-12-31): 1.4e24 FLOP, confidence ConfidentK-EXAONE (LG AI Research, 2025-12-31): 1.5e24 FLOP, confidence ConfidentKimi K2.5 (Moonshot, 2026-02-02): 5.8e24 FLOP, confidence LikelyGLM-5 (Z.ai (Zhipu AI), 2026-02-12): 6.8e24 FLOP, confidence LikelyComposer 2 (Cursor, 2026-03-19): 2.3e25 FLOP, confidence ConfidentMolmoAct 2 (Allen Institute for AI,University of Washington,National University of Singapore,University of Pennsylvania,Johns Hopkins University,Amazon,Cortex AI, 2026-04-05): 1.1e22 FLOP, confidence LikelyEXAONE 4.5 (LG AI Research, 2026-04-09): 3.9e24 FLOP, confidence ConfidentMiMo-V2.5-Pro (Xiaomi Corp, 2026-04-23): 6.8e24 FLOP, confidence LikelyDeepSeek-V4-Pro (DeepSeek, 2026-04-24): 9.7e24 FLOP, confidence LikelyDeepSeek-V4-Flash (DeepSeek, 2026-04-24): 2.5e24 FLOP, confidence LikelyComposer 2.5 (Cursor, 2026-05-18): 3.9e25 FLOP, confidence LikelyNemotron 3 Ultra (Nvidia, 2026-06-04): 6.6e24 FLOP, confidence LikelySolar Open2 250B (Upstage, 2026-06-28): 1.1e24 FLOP, confidence LikelyLongCat-2.0 (Meituan Inc, 2026-06-29): 1.0e25 FLOP, confidence LikelyInkling (Thinking Machines, 2026-07-15): 1.1e25 FLOP, confidence LikelyKimi K3 (Moonshot, 2026-07-16): 2.0e25 FLOP, confidence SpeculativeA.X K2 (SK Telecom, 2026-07-29): 1.8e24 FLOP, confidence ConfidentK-EXAONE 2.0 (LG AI Research, 2026-07-31): 3.5e24 FLOP, confidence ConfidentMotif-3 (Motif Technologies, 2026-08-07): 1.1e24 FLOP, confidence ConfidentGLM-5.3-Flash (Z.ai (Zhipu AI), 2026-08-20): 3.2e24 FLOP, confidence LikelyWeatherNext 3 (Google DeepMind,Google Research, 2026-09-03): 2.1e23 FLOP, confidence SpeculativeGPT-6 Astra (OpenAI, 2026-09-03): 9.4e26 FLOP, confidence LikelyDeepSeek V4.1 Flash (DeepSeek, 2026-09-09): 4.3e24 FLOP, confidence LikelyPublication yearTraining compute, FLOP
Reported model Largest run to date EU systemic-risk threshold 468 models published 2012 onward with a compute figure. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , today Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/capability/.

Compute figures are frequently estimated rather than disclosed, and Epoch marks its own confidence on each. Faded points are the ones marked speculative. Models whose compute is unknown are absent entirely, which biases the picture toward labs that publish. Treat anything near the threshold line as indicative, never as a legal finding about a particular model. Epoch's own interactive version of this chart is at epoch.ai/data/notable-ai-models.

What the training runs cost

Disclosed or estimated cost of the final training run, in 2023 dollars, for the 178 models where a figure exists. Also logarithmic.

$1 $10 $100 $1k $10k $100k $1m $10m $100m $1.0bn 2012 2014 2016 2018 2020 2022 2024 2026 Dropout (ImageNet) (University of Toronto, 2012-06-03): $8 Unsupervised High-level Feature Learner (Google, 2012-07-12): $16 AlexNet (University of Toronto, 2012-09-30): $16 DNN EM segmentation (IDSIA,SUPSI, 2012-12-03): $4 DistBelief NNLM (Google, 2013-01-16): $3k Visualizing CNNs (New York University (NYU), 2013-11-12): $13 TransE (Universite de Technologie de Compiègne - CNRS,Google, 2013-12-05): $30 SPPNet (Microsoft,Xi’an Jiaotong University,University of Science and Technology of China (USTC), 2014-06-18): $52 VGG16 (University of Oxford, 2014-09-04): $239 MSRA (C, PReLU) (Microsoft Research, 2015-02-06): $1k AlphaGo Fan (DeepMind, 2015-10-01): $5k Inception v3 (Google,University College London (UCL), 2015-12-02): $1k AlphaGo Lee (DeepMind, 2016-01-27): $22k GNMT (Google, 2016-09-26): $201k Xception (Google, 2016-10-07): $13k NASv3 (CIFAR-10) (Google Brain, 2016-11-05): $21k BIDAF (University of Washington,Allen Institute for AI, 2016-11-05): $41 PolyNet (Chinese University of Hong Kong (CUHK), 2016-11-17): $617 MoE-Multi (Jagiellonian University,Google Brain, 2017-01-23): $4k Transformer (Google Research,Google Brain, 2017-06-12): $438 JFT (Google Research,Carnegie Mellon University (CMU), 2017-07-10): $18k AlphaGo Zero (DeepMind, 2017-10-18): $1k AlphaGo Master (DeepMind, 2017-10-19): $471k AlphaZero (DeepMind, 2017-12-05): $230k IMPALA (DeepMind, 2018-02-05): $53 ResNeXt-101 32x48d (Facebook, 2018-05-02): $134k Big Transformer for Back-Translation (Facebook AI Research,Google Brain, 2018-08-28): $2k Transformer + Simple Recurrent Unit (ASAPP,Cornell University,Google,Princeton University, 2018-09-17): $45 Transformer (Adaptive Input Embeddings) WT103 (Facebook AI Research, 2018-09-28): $3k BERT-Large (Google, 2018-10-11): $2k Mesh-TensorFlow Transformer 4.9B (language) (Google Brain, 2018-11-05): $935 Mesh-TensorFlow Transformer 2.9B (translation) (Google Brain, 2018-11-05): $396 GPT-2 (1.5B) (OpenAI, 2019-02-14): $4k KataGo (Jane Street, 2019-02-27): $105 SciBERT (Allen Institute for AI, 2019-03-26): $247 XLNet (Carnegie Mellon University (CMU),Google Brain, 2019-06-01): $14k RoBERTa Large (Facebook,University of Washington, 2019-07-01): $85k Megatron-LM (8.3B) (Nvidia, 2019-09-17): $109k Megatron-BERT (Nvidia, 2019-09-17): $172k T5-11B (Google, 2019-10-23): $78k T5-3B (Google, 2019-10-23): $17k AlphaStar (DeepMind, 2019-10-30): $131k XLM-RoBERTa (Facebook AI, 2019-11-05): $78k CamemBERT (Facebook,INRIA,Sorbonne University, 2019-11-10): $2k Noisy Student (L2) (Carnegie Mellon University (CMU),Google, 2019-11-11): $46k OpenAI Five Rerun (OpenAI, 2019-12-13): $321k OpenAI Five (OpenAI, 2019-12-13): $4m DD-PPO (Georgia Institute of Technology,Facebook AI Research,Oregon State University,Simon Fraser University, 2019-12-19): $2k ContextNet + Noisy Student (Google, 2020-01-19): $14k Meena (Google Brain, 2020-01-28): $215k ALBERT-xxlarge (Toyota Technological Institute at Chicago,Google, 2020-02-09): $4k Turing-NLG (Microsoft, 2020-02-13): $52k Once for All (MIT-IBM Watson AI Lab,Massachusetts Institute of Technology (MIT),IBM, 2020-04-29): $2k UnifiedQA (Allen Institute for AI,University of Washington, 2020-05-02): $59 DETR (Facebook, 2020-05-26): $960 GPT-3 175B (davinci) (OpenAI, 2020-05-28): $2m GShard (dense) (Google, 2020-06-30): $266k ProBERTa (University of Illinois Urbana-Champaign (UIUC),Reed College, 2020-09-01): $26 LUKE (University of Washington,National Institute of Informatics, 2020-10-02): $4k mT5-XXL (Google,Google Research, 2020-10-20): $179k Conformer + Wav2vec 2.0 + Noisy Student (Google,Google Research,Google Brain, 2020-10-20): $9k German ELECTRA Large (deepset,Bayerische Staatsbibliothek Muenchen, 2020-10-21): $2k wave2vec 2.0 LARGE (Facebook, 2020-10-22): $5k ViT-Huge/14 (Google Brain,Google Research, 2020-10-22): $7k AlphaFold 2 (DeepMind, 2020-11-30): $4k CPM-Large (Tsinghua University,Beijing Academy of Artificial Intelligence / BAAI, 2020-12-01): $7k ESM1b (Facebook AI Research,New York University (NYU), 2020-12-15): $924 DALL-E (OpenAI, 2021-01-05): $126k CLIP (ViT L/14@336px) (OpenAI, 2021-01-05): $25k Switch (Google, 2021-01-11): $145k MSA Transformer (Facebook AI Research,University of California (UC) Berkeley,New York University (NYU), 2021-02-13): $13k Meta Pseudo Labels (Google Brain,Google AI, 2021-03-01): $54k M6-T (Alibaba, 2021-03-05): $13k PLUG (Alibaba, 2021-04-19): $109k ProtBERT-BFD (Technical University of Munich,Nvidia,Seoul National University,Google,Oak Ridge National Laboratory,Med AI Technology, 2021-05-04): $45k ADM (OpenAI, 2021-05-11): $11k MedBERT (Peng Cheng Laboratory,University of Texas at Houston, 2021-05-20): $62 CogView (Tsinghua University,Alibaba DAMO Academy, 2021-05-26): $60k ByT5-XXL (Google,Google Research, 2021-05-28): $92k ViT-G/14 (Google Brain,Google Research, 2021-06-08): $4k EMDR (Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),McGill University,DeepMind, 2021-06-09): $3k CoAtNet (Google,Google Research,Google Brain, 2021-06-09): $2k DeBERTa (Microsoft, 2021-06-10): $7k Denoising Diffusion Probabilistic Models (LSUN Bedroom) (University of California (UC) Berkeley, 2021-06-11): $436 ALIGN (Google Research, 2021-06-11): $33k EfficientNetV2-XL (Google,Google Brain, 2021-06-23): $104 ERNIE 3.0 (Baidu, 2021-07-05): $39k GOAT (DeepMind, 2021-07-27): $85k SEER (Facebook AI Research,INRIA, 2021-07-29): $34k Jurassic-1-Jumbo (AI21 Labs, 2021-08-11): $837k HyperCLOVA 204B (NAVER, 2021-09-10): $442k AlphaFold-Multimer (Google DeepMind,DeepMind, 2021-10-04): $8k Megatron-Turing NLG 530B (Microsoft,Nvidia, 2021-10-11): $4m Yuan 1.0 (Inspur, 2021-10-12): $592k CodeT5-base (Salesforce,Nanyang Technological University, 2021-11-01): $3k Swin Transformer V2 (SwinV2-G) (Microsoft Research Asia, 2021-11-18): $2k BASIC-L (Google, 2021-11-19): $2k Florence (Microsoft, 2021-11-22): $107k Gopher (280B) (DeepMind, 2021-12-08): $641k GLaM (Google, 2021-12-13): $541k XGLM-7.5B (Meta AI,Facebook AI Research, 2021-12-20): $104k Detic (Meta AI,University of Texas at Austin, 2022-01-07): $191 GPT-NeoX-20B (EleutherAI, 2022-02-09): $184k LaMDA (Google, 2022-02-10): $230k GPT-3.5 (davinci-002) (OpenAI, 2022-03-15): $5m PaLM (540B) (Google Research, 2022-04-04): $3m Stable Diffusion (LDM-KL-8-G) (Runway,Ludwig Maximilian University of Munich,Heidelberg University, 2022-04-13): $111k Flamingo (DeepMind, 2022-04-29): $183k OPT-175B (Meta AI, 2022-05-02): $734k UL2 (Google Research,Google Brain, 2022-05-10): $127k Gato (DeepMind, 2022-05-12): $4k Imagen (Google Brain, 2022-05-23): $8k Tranception (University of Oxford,Harvard Medical School,Cohere, 2022-05-27): $15k CoCa (Google Research, 2022-06-14): $78k Parti (Google Research, 2022-06-22): $427k ProGen2-xlarge (Salesforce Research,Columbia University,Johns Hopkins University, 2022-06-27): $12k CodeT5-large (Salesforce, 2022-07-05): $4k NLLB (Meta AI, 2022-07-06): $51k BLOOM-176B (Hugging Face,BigScience, 2022-07-11): $996k ESM2-15B (Meta AI,New York University (NYU),Stanford University,Massachusetts Institute of Technology (MIT), 2022-07-21): $163k OmegaPLM (Massachusetts Institute of Technology (MIT),Westlake University, 2022-07-22): $52k AlexaTM 20B (Amazon, 2022-08-02): $268k GLM-130B (Tsinghua University, 2022-08-04): $820k PaLI (Google, 2022-09-14): $51k EVA-01 (Beijing Academy of Artificial Intelligence / BAAI,Huazhong University of Science and Technology,Zhejiang University (ZJU),Beijing Institute of Technology, 2022-11-14): $29k Galactica (Meta AI, 2022-11-16): $591k Fusion in Encoder (Samsung, 2022-11-18): $233 AR-LDM (Alibaba,University of Waterloo,Vector Institute, 2022-11-20): $746 Discriminator Guidance (Korea Advanced Institute of Science and Technology (KAIST),NAVER, 2022-11-28): $338 VALL-E (Microsoft, 2023-01-05): $11 Nucleotide Transformer (Nvidia,Technical University of Munich,InstaDeep, 2023-01-15): $51k Ankh_large (Technical University of Munich,Columbia University, 2023-01-16): $5k DDPM-IP (CelebA) (Utrecht University, 2023-01-27): $390 BLIP-2 (Q-Former) (Salesforce Research, 2023-01-30): $2k ViT-22B (Google, 2023-02-10): $286k LLaMA-65B (Meta AI, 2023-02-24): $578k DiT-XL/2 (New York University (NYU),University of California (UC) Berkeley, 2023-03-02): $111k AudioGen (Meta AI,Hebrew University of Jerusalem, 2023-03-05): $9k GPT-4 (Mar 2023) (OpenAI, 2023-03-15): $37m Falcon-40B (Technology Innovation Institute, 2023-03-15): $320k VideoMAE V2 (Nanjing University,Shenzhen Institute of Advanced Technology,Shanghai AI Lab, 2023-03-29): $18k BloombergGPT (Bloomberg,Johns Hopkins University, 2023-03-30): $370k Segment Anything Model (Meta AI, 2023-04-05): $16k Incoder-6.7B (Facebook AI Research,University of Washington,University of California (UC) Berkeley,Carnegie Mellon University (CMU),Toyota Technological Institute at Chicago, 2023-04-09): $3k DINOv2 (Facebook AI Research,INRIA, 2023-04-14): $10k LLaVA (University of Wisconsin Madison,Microsoft Research,Columbia University, 2023-04-17): $42 StarCoder (Hugging Face,ServiceNow,Northeastern University,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),Carnegie Mellon University (CMU),Johns Hopkins University,Leipzig University,ScaDS.AI,Queen Mary University of London,Roblox,Sea AI Lab,Technion - Israel Institute of Technology,Monash University,CSIRO,Data61,McGill University,Saama,University of British Columbia (UBC),Massachusetts Institute of Technology (MIT),Technical University of Munich,IBM,University of Vermont,UnfoldML,SAP,University of Notre Dame,Columbia University,New York University (NYU),University of Allahabad,Discover Dollar,Toloka,Telefonica,Stanford University,Weizmann Institute of Science,Alan Turing Institute,Wellesley College,EleutherAI,Forschungszentrum Julich, 2023-05-09): $212k PaLM 2 (Google, 2023-05-10): $5m HyenaDNA (Stanford University,Harvard University,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),University of Montreal / Université de Montréal, 2023-06-27): $5k Pangu-Weather (Huawei, 2023-07-05): $51k xTrimoPGLM -100B (Tsinghua University,BioMap Research, 2023-07-06): $2m InternLM (Shanghai AI Lab,SenseTime, 2023-07-06): $2m Claude 2 (Anthropic, 2023-07-11): $5m Llama 2-70B (Meta AI, 2023-07-18): $1m Llama 2-7B (Meta AI, 2023-07-18): $114k Falcon-180B (Technology Innovation Institute, 2023-09-06): $11m Amazon Titan (Amazon, 2023-09-28): $8m CODEFUSION (Python) (Microsoft,Microsoft Research, 2023-10-26): $9 MultiBand Diffusion (Meta AI,Hebrew University of Jerusalem,LORIA, 2023-11-08): $23 SPHINX (Llama 2 13B) (Shanghai AI Lab,Chinese University of Hong Kong (CUHK),ShanghaiTech University, 2023-11-13): $239k Nemotron-3-8B (Nvidia, 2023-11-15): $214k Inflection-2 (Inflection AI, 2023-11-22): $13m Gemini 1.0 Ultra (Google DeepMind, 2023-12-06): $31m MegaScale (Production) (ByteDance,Peking University, 2024-02-23): $3m Inflection-2.5 (Inflection AI, 2024-03-07): $12m Nemotron-4 340B (Nvidia, 2024-06-14): $21m Claude 3.5 Sonnet (Anthropic, 2024-06-20): $26m Llama 3.1-405B (Meta AI, 2024-07-23): $53m Grok-2 (xAI, 2024-08-13): $32m DeepSeek-V3 (DeepSeek, 2024-12-24): $5m DeepSeek-R1 (DeepSeek, 2025-01-20): $7m Grok 3 (xAI, 2025-02-17): $218m GPT-4.5 (OpenAI, 2025-02-27): $366m DeepSeek-V3 (Mar 2025) (DeepSeek, 2025-03-24): $5m Llama 4 Behemoth (preview) (Meta AI, 2025-04-05): $45m DeepSeek-R1 (May 2025) (DeepSeek, 2025-05-28): $7m Grok 4 (xAI, 2025-07-09): $388m A.X K2 (SK Telecom, 2026-07-29): $12m Publication year

Cost is known for 178 of 540 models with a compute figure, so this is a sample of the disclosed, not a census of the spent. It counts the final training run only: not the failed runs before it, not the staff, not the data, not the cluster sitting idle between runs. The most expensive here is Grok 4 at $388m. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , today Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/capability/.

How models ship

Notable models by publication year and how they were made available: Open weights A release in which the trained parameters are distributed and can be run by a third party. It does not imply an open-source licence, and the licence attached to a given release may restrict use. Glossary entry , API access A release in which a model can be queried over a network but its weights are not distributed. One of three release types recorded for notable models. Glossary entry , or never released.

17
18
19
20
21
22
23
24
25
26
Open weights API only Unreleased Other Not recorded Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , today Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/capability/.
Year Open weightsAPI onlyUnreleasedOtherNot recorded Total
2026 295343 89
2025 45557 107
2024 394316 98
2023 592436 119
2022 4113361 91
2021 41929 79
2020 28121 50
2019 3137 68
2018 91207 37
2017 112429 55

The current year is partial and will fill in. "Not recorded" is its own column on purpose: folding it into "unreleased" would turn a gap in the record into a claim that a model was never released, which is a much stronger statement than the data supports.

How far behind the open-weight frontier is

Largest training run to date over all models, against the same over models released with open weights. Both are Running maximum The highest value recorded up to each date. It rises or holds level and never falls, so a line fitted through one cannot slope downward. Glossary entry , so neither line can fall.

  • 8 months between the overall frontier reaching 3.8e25 FLOP and an open-weight model doing so Measured horizontally: the date the open-weight record was set, minus the date the first model of any kind reached that compute level. Computed by this site from Epoch's figures. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , today Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/capability/. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , today Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/capability/.
  • 27 months since the open-weight record last moved, set by Llama 3.1-405B
  • 25x more compute in GPT-6 Astra than in the largest open-weight model
  • 15 times the open-weight record has been broken since 2015
Largest training run to date, all models against open-weight models Running maximum of training compute by publication date. Accessibility is Epoch's classification; models with no accessibility recorded are counted in the overall line and not in the open-weight one. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , today Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/capability/. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , today Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/capability/. 2012 to 2027
1e171e191e211e231e251e272016201820202022202420268 months behind
all models open weights
Use this figure

An SVG image of this figure, with its sources, date and address printed in it, linked to this section. SVG file. All embeddable figures. 

<a href="https://aisafetytracker.org/capability/#open-weights"><img src="https://aisafetytracker.org/embed/open-weight-lag.svg" alt="Step chart of the largest training run to date for all AI models against open-weight models, with the lag between them in months, from AI Safety Tracker" width="948" height="502" loading="lazy" style="max-width:100%;height:auto"></a>

Both lines are running maxima, so a flat stretch means no larger run has been reported, not that anything got smaller. The open-weight line counts only models Epoch records as released with open weights and with a compute figure attached; a model with unrecorded accessibility falls into the overall line and not this one, so the true open-weight frontier can only be at or above what is drawn. Compute figures are frequently estimates, and they arrive late: recent models with a compute figure attached are almost all open-weight ones, because closed developers do not publish and an outside estimate takes time, so the right-hand end of the overall line is a lower bound too. Training compute is also not capability: a smaller, better-trained model can beat a larger one, which is the main reason this gap has narrowed in benchmark terms faster than it has in FLOP.

Who builds them

Notable models published since 2023, by organisation and by the country of the organisation that built them.

By organisation 147 in total
Organisation Models
OpenAI 58
Google DeepMind 36
Alibaba 31
Anthropic 22
Meta AI 21
ByteDance 10
DeepSeek 10
Nvidia 9
xAI 9
Moonshot 8
Show the remaining 137
Z.ai (Zhipu AI) 8
LG AI Research 7
Mistral AI 6
Tencent 6
Google 5
Apple 4
Baidu 4
01.AI 3
Character.ai 3
Google Research 3
Microsoft 3
MiniMax 3
Amazon 2
Cursor 2
Huawei 2
Inception Labs 2
Inflection AI 2
Kunlun Inc. 2
Meituan Inc 2
National University of Singapore 2
Nvidia,Massachusetts Institute of Technology (MIT) 2
Runway 2
SK Telecom 2
Stability AI 2
Technology Innovation Institute 2
Thinking Machines 2
Tsinghua University 2
Upstage 2
Writer 2
Xiaomi Corp 2
Z.ai (Zhipu AI),Tsinghua University 2
Adobe 1
AI21 Labs 1
Alibaba,Huazhong University of Science and Technology 1
Allen Institute for AI 1
Allen Institute for AI,University of Washington,National University of Singapore,University of Pennsylvania,Johns Hopkins University,Amazon,Cortex AI 1
Ant Group 1
Bloomberg,Johns Hopkins University 1
ByteDance,Nanyang Technological University,Chinese University of Hong Kong (CUHK),Hong Kong University of Science and Technology (HKUST) 1
ByteDance,Peking University 1
Carnegie Mellon University (CMU) 1
Carnegie Mellon University (CMU),Princeton University 1
Cerebras Systems,Mohamed bin Zayed University of Artificial Intelligence (MBZUAI),Inception G42 1
Chai discovery 1
China Telecom 1
Cognition 1
Cohere 1
Cohere Labs (formerly Cohere for AI),Brown University,Cohere,Carnegie Mellon University (CMU),Massachusetts Institute of Technology (MIT) 1
Columbia University,Apple 1
Databricks 1
DeepL 1
DeepMind,University of Toronto 1
ETH Zurich,Disney Research 1
EvolutionaryScale,University of California (UC) Berkeley 1
Facebook AI Research,INRIA 1
Facebook AI Research,University of Washington,University of California (UC) Berkeley,Carnegie Mellon University (CMU),Toyota Technological Institute at Chicago 1
Facebook,INRIA,University of California (UC) Berkeley 1
Generalist 1
Google DeepMind,Google Research 1
Google DeepMind,Google Research,Google Cloud 1
Google DeepMind,Isomorphic Labs 1
Google DeepMind,New York University (NYU) 1
Google DeepMind,University of California San Diego,Stanford University 1
Google Research,DeepMind 1
Google,Google DeepMind 1
Google,TU Berlin 1
Google,University of California Los Angeles (UCLA) 1
Harrison.ai 1
Huawei Noah's Ark Lab 1
Hugging Face,ServiceNow,Northeastern University,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),Carnegie Mellon University (CMU),Johns Hopkins University,Leipzig University,ScaDS.AI,Queen Mary University of London,Roblox,Sea AI Lab,Technion - Israel Institute of Technology,Monash University,CSIRO,Data61,McGill University,Saama,University of British Columbia (UBC),Massachusetts Institute of Technology (MIT),Technical University of Munich,IBM,University of Vermont,UnfoldML,SAP,University of Notre Dame,Columbia University,New York University (NYU),University of Allahabad,Discover Dollar,Toloka,Telefonica,Stanford University,Weizmann Institute of Science,Alan Turing Institute,Wellesley College,EleutherAI,Forschungszentrum Julich 1
Inspur 1
Intel Labs 1
King Abdullah University of Science and Technology (KAUST),Karolinska Institute 1
Korea University,Korea Advanced Institute of Science and Technology (KAIST),LG 1
Magic 1
Meta AI,Hebrew University of Jerusalem 1
Meta AI,Hebrew University of Jerusalem,LORIA 1
Meta AI,Stanford University 1
Microsoft,Microsoft Research 1
Motif Technologies 1
Nanjing University,Shenzhen Institute of Advanced Technology,Shanghai AI Lab 1
NAVER 1
NC AI 1
New York University (NYU) 1
New York University (NYU),University of California (UC) Berkeley 1
Nvidia,Massachusetts Institute of Technology (MIT),University of California (UC) Berkeley,University of California San Diego,University of Washington,Tsinghua University 1
Nvidia,Meta AI 1
Nvidia,Nanjing University,Tsinghua University,Hong Kong Polytechnic University,Johns Hopkins University,New York University (NYU) 1
Nvidia,Technical University of Munich,InstaDeep 1
Nvidia,University of Toronto,Vector Institute,University of Illinois Urbana-Champaign (UIUC) 1
OpenEvidence 1
Oppo Mobile Telecommunications 1
Perplexity 1
Physical Intelligence 1
Prime Intellect 1
Reka AI 1
rinna 1
Salesforce 1
Salesforce Research 1
Salesforce Research,Hong Kong University of Science and Technology (HKUST),Nanyang Technological University 1
Sapient Intelligence 1
Saudi Aramco 1
Saudi Data and Artificial Intelligence Authority 1
SenseTime 1
Shanghai AI Lab 1
Shanghai AI Lab,Chinese University of Hong Kong (CUHK),ShanghaiTech University 1
Shanghai AI Lab,SenseTime 1
Shanghai Qi Zhi institute,Stanford University,Carnegie Mellon University (CMU),Tsinghua University 1
Stability AI,University of Illinois Urbana-Champaign (UIUC) 1
Stanford University,CZ Biohub Network 1
Stanford University,Google Research,Georgia Institute of Technology 1
Stanford University,Harvard University,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms),University of Montreal / Université de Montréal 1
Stanford University,Princeton University 1
Stanford University,Sony 1
Stanford University,University of California (UC) Berkeley,Toyota Research Institute,Google DeepMind,Massachusetts Institute of Technology (MIT),Physical Intelligence 1
StepFun 1
Suno 1
Technical University of Munich,Columbia University 1
TensorTour 1
Tsinghua University,BioMap Research 1
Tsinghua University,Nanyang Technological University,Carnegie Mellon University (CMU) 1
Tsinghua University,Tencent,Nanyang Technological University 1
Tsinghua University,University of Illinois Urbana-Champaign (UIUC),Shanghai AI Lab,Peking University,Shanghai Jiao Tong University,CUHK Shenzhen Research Institute 1
Tsinghua University,Z.ai (Zhipu AI) 1
Tsinghua University,Z.ai (Zhipu AI),Beihang University 1
University of California (UC) Berkeley,California Institute of Technology,University of Toronto,University of Wisconsin Madison,Texas A&M,Nvidia,Mila - Quebec AI (originally Montreal Institute for Learning Algorithms) 1
University of California (UC) Berkeley,Stanford University,Carnegie Mellon University (CMU),DeepMind 1
University of California Los Angeles (UCLA),Columbia University,New York University (NYU) 1
University of Minnesota,Grammarly 1
University of Southern California,Georgia Institute of Technology,Stanford University,Nvidia 1
University of Turku,Hugging Face 1
University of Washington,Seoul National University,University of Sheffield 1
University of Wisconsin Madison,Microsoft Research 1
University of Wisconsin Madison,Microsoft Research,Columbia University 1
Utrecht University 1
Westlake University,Tsinghua University,Toyota Technological Institute at Chicago 1
Zhuiyi Technology 1
By country of organisation 17 in total
Country Models
United States of America 251
China 117
South Korea 15
France 6
United Kingdom 4
Saudi Arabia 3
Singapore 3
Canada 2
Germany 2
United Arab Emirates 2
Show the remaining 7
Australia 1
Finland 1
Hong Kong 1
Israel 1
Japan 1
Netherlands 1
Switzerland 1

413 models published since 2023, of which 412 have a country recorded. Country is the country of the organisation, not of the compute, the data or the people. A multinational lab is counted once, at its listed home. Inclusion in Epoch's database is editorial, so these are counts of what has been judged notable rather than of everything built. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , today Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/capability/.

The compute thresholds written into law

Every Compute threshold A quantity of training compute written into a statute, executive action or official policy document, above which additional obligations apply. Recorded here with the instrument that sets it and the number of models above it. Glossary entry set by a statute, an executive action or an official policy document, the day each took its final form, and how many models with a published compute figure sat above it that day and sit above it now.

  • 4 of 6 thresholds written after a model had already crossed them Compute Threshold Index This site Licence MIT Updated quarterly Retrieved , 5 days ago Data JSON CSV Training-compute thresholds written into law, executive action or published policy, with the text that sets each one. Whether a given model crosses a threshold here is arithmetic on Epoch's compute estimates, most of which are estimates rather than disclosures, and is not a legal finding about any model or developer. Cite AI Safety Tracker, "Compute Threshold Index". https://aisafetytracker.org/capability/#thresholds. Retrieved 23 October 2026. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , 5 days ago Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 23 October 2026 via AI Safety Tracker, https://aisafetytracker.org/capability/.
  • 5 models already above 1e25 when the EU AI Act was signed, 13 Jun 2024
  • 7 months from the first model above 1e26 to the first 1e26 threshold in force
1e241e251e261e2720232024202520262027EO 14110, 2023-10EU AI Act, 2024-06SB 1047, 2024-09SB 53, 2025-09RAISE Act, 2025-12Korean decree, 2026-01
Reported model Largest run to date In force Enacted, not yet applicable Revoked, or a vetoed bill
Use this figure

An SVG image of this figure, with its sources, date and address printed in it, linked to this section. SVG file. All embeddable figures. 

<a href="https://aisafetytracker.org/capability/#thresholds"><img src="https://aisafetytracker.org/embed/compute-thresholds.svg" alt="Timeline of the training-compute thresholds written into law and policy against the AI models with a published compute figure, from AI Safety Tracker" width="948" height="485" loading="lazy" style="max-width:100%;height:auto"></a>
Jurisdiction Instrument Threshold Status Dated Above that day First crossed Above now
European Union Regulation (EU) 2024/1689, Article 51(2) 1e25 in force 2024-06-13 signed 5 GPT-4 (Mar 2023), 2023-03-15 24
United States (California) SB 53, Transparency in Frontier Artificial Intelligence Act 1e26 in force 2025-09-29 signed 3 Grok 3, 2025-02-17 4
South Korea Enforcement Decree of the AI Basic Act, Article 23 1e26 in force 2026-01-20 approved 3 Grok 3, 2025-02-17 4
United States (federal) Executive Order 14110, section 4.2 1e26 revoked 2023-10-30 signed 0 Grok 3, 2025-02-17 4
United States (California) SB 1047, Safe and Secure Innovation for Frontier AI Models Act 1e26 vetoed 2024-09-29 vetoed 0 Grok 3, 2025-02-17 4
United States (New York) RAISE Act, Responsible AI Safety and Education Act 1e26 enacted 2025-12-19 signed 3 Grok 3, 2025-02-17 4
What each threshold actually requires 6

  • Regulation (EU) 2024/1689, Article 51(2) 1e25 in force

    Applies to General-purpose AI models placed on the EU market

    Consequence Presumption of systemic risk, bringing model evaluation, adversarial testing, incident reporting and cybersecurity obligations.

    A rebuttable presumption rather than a hard line: a developer can argue its model does not carry systemic risk despite the compute, and the Commission can designate a model that falls below it.

  • SB 53, Transparency in Frontier Artificial Intelligence Act 1e26 in force

    Applies to Frontier developers training models above the threshold

    Consequence Duty to publish a frontier AI framework and transparency report, and to report critical safety incidents to the state.

    Counts the initial training run plus later fine-tuning and material modification, so it is a cumulative figure rather than a single-run one. Larger obligations attach above an annual revenue threshold as well.

  • Enforcement Decree of the AI Basic Act, Article 23 1e26 in force

    Applies to AI systems trained with cumulative compute at or above the threshold that also apply state-of-the-art technology and pose wide and serious risk to fundamental rights

    Consequence Duty to identify, assess and mitigate risks across the life cycle, run a safety-incident response system, and submit the results to the Ministry of Science and ICT.

    Cumulative training compute, and the first of three conditions that must all hold, so crossing 1e26 alone does not bring a model into scope. In force from 22 January 2026 under the AI Basic Act, with fines deferred for a grace period of at least a year.

  • Executive Order 14110, section 4.2 1e26 revoked

    Applies to Developers of dual-use foundation models

    Consequence Reporting of training runs, results of red-team testing and security measures.

    The first compute threshold in any official instrument, revoked in January 2025. Included because it is still widely quoted as current, and because California later adopted the same number.

  • SB 1047, Safe and Secure Innovation for Frontier AI Models Act 1e26 vetoed

    Applies to Developers of covered models above the threshold and above $100m training cost

    Consequence Would have required a safety protocol, third-party audit and a shutdown capability.

    Vetoed in September 2024. Its successor, SB 53, kept the compute threshold and dropped the audit and shutdown requirements.

  • RAISE Act, Responsible AI Safety and Education Act 1e26 enacted applies from 2027-01-01

    Applies to Frontier developers training models above the threshold, with the heavier duties on those with over $500 million in annual revenue

    Consequence Duty to publish safety and security protocols and to report safety incidents to the state.

    Signed in December 2025 and amended in March 2026 to track California's SB 53. It applies from 1 January 2027, so it is counted here as enacted rather than in force.

Six thresholds exist, five of them at 1e26, three in force, one enacted but not yet applicable, one revoked and one vetoed. Whether a model sits above a line is arithmetic on Epoch's compute figures, most of which are estimates rather than disclosures, and the EU threshold is a rebuttable presumption rather than a hard boundary. Nothing here is a legal finding about any model or developer. Thresholds are our own index of primary instruments. Model compute figures are Epoch's. Compute Threshold Index This site Licence MIT Updated quarterly Retrieved , 5 days ago Data JSON CSV Training-compute thresholds written into law, executive action or published policy, with the text that sets each one. Whether a given model crosses a threshold here is arithmetic on Epoch's compute estimates, most of which are estimates rather than disclosures, and is not a legal finding about any model or developer. Cite AI Safety Tracker, "Compute Threshold Index". https://aisafetytracker.org/capability/#thresholds. Retrieved 23 October 2026. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , 5 days ago Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 23 October 2026 via AI Safety Tracker, https://aisafetytracker.org/capability/.

Benchmark scores, difficulty tiers and the rate at which tests are being solved have moved to Progress.

Related

  • Progress Benchmark scores for the models these training runs produced.
  • Datacentres The sites that supply training compute, by power and owner.
  • Law and policy Eleven jurisdictions' instruments, among them those that set the thresholds.