AI in charts

Training compute since 1950, who builds the models, and what today's models cost, how fast they run and how they score. Built from Epoch AI's records and the index's own data, updated Nov 12, 2026.

The compute used to train AI has grown about 4.3× a year since 2010

Epoch AI tracks 1,102 notable AI models. For 546 of them it has an estimate of the training compute: the number of calculations used to train the model, counted in FLOP.

Before 2010, training compute grew about 1.5× a year, doubling roughly every 20 months.

Since 2010 it has grown about 4.3× a year, doubling roughly every six months. Each step up the scale is ten times more compute.

The largest training run on record is GPT-6 Astra (OpenAI, Sep 2026), at an estimated 9.4 × 10²⁶ FLOP.

Data: Epoch AI, CC BY 4.0.

Show the data as a table
ModelOrganizationPublishedTraining compute (FLOP)
GPT-6 AstraOpenAISep 3, 20269.4 × 10²⁶
Grok 4xAIJul 9, 20255 × 10²⁶
GPT-4.5OpenAIFeb 27, 20253.8 × 10²⁶
Grok 3xAIFeb 17, 20253.5 × 10²⁶
Gemini 1.0 UltraGoogle DeepMindDec 6, 20235 × 10²⁵
GPT-4 (Mar 2023)OpenAIMar 15, 20232.1 × 10²⁵
Minerva (540B)GoogleJun 29, 20222.7 × 10²⁴
GPT-3.5 (davinci-002)OpenAIMar 15, 20222.6 × 10²⁴
FLAN 137BGoogle ResearchSep 3, 20212 × 10²⁴
Jurassic-1-JumboAI21 LabsAug 11, 20213.7 × 10²³
GPT-3 175B (davinci)OpenAIMay 28, 20203.1 × 10²³
MeenaGoogle BrainJan 28, 20201.1 × 10²³
AlphaStarDeepMindOct 30, 20191.1 × 10²³
T5-11BGoogleOct 23, 20193.3 × 10²²
Megatron-BERTNvidiaSep 17, 20192.2 × 10²²
Megatron-LM (8.3B)NvidiaSep 17, 20199.1 × 10²¹
ResNeXt-101 32x48dFacebookMay 2, 20188.7 × 10²¹
GNMTGoogleSep 26, 20166.6 × 10²¹
AlphaGo LeeDeepMindJan 27, 20161.9 × 10²¹
AlphaGo FanDeepMindOct 1, 20153.8 × 10²⁰
SNM-skipGoogleDec 3, 20143 × 10²⁰
Seq2Seq LSTMGoogleSep 10, 20145.6 × 10¹⁹
VGG16University of OxfordSep 4, 20141.2 × 10¹⁹
VGG19University of OxfordSep 4, 20141.1 × 10¹⁹
SPPNetMicrosoft, Xi’an Jiaotong UniversityJun 18, 20143.4 × 10¹⁸
DistBelief NNLMGoogleJan 16, 20132.6 × 10¹⁸
SB-LMGoogleJun 22, 20071.4 × 10¹⁸
Neural LMUniversity of Montreal / Université de MontréalNov 28, 20006.3 × 10¹⁵
PoE MNISTUniversity College London (UCL)Nov 28, 20005.2 × 10¹³
LSTMTechnical University of MunichNov 15, 19973.2 × 10¹³
Predictive Coding NNTechnical University of MunichDec 2, 19941.9 × 10¹³
TD-GammonIBMMay 1, 19921.8 × 10¹³
Zip CNNAT&T, Bell LaboratoriesDec 1, 19891.5 × 10¹²
Handwritten digit recognition networkAT&TNov 27, 19891.8 × 10¹¹
NetTalk (transcription)Princeton UniversityJun 6, 19872.8 × 10¹⁰
Perceptron (1960)Cornell Aeronautical LaboratoryMar 30, 19607.2 × 10⁸
Pandemonium (morse)Massachusetts Institute of Technology (MIT)Feb 1, 19596 × 10⁸
Perceptron Mark ICornell Aeronautical Laboratory, Cornell UniversityJan 1, 19576.9 × 10⁵
TheseusBell LaboratoriesJul 2, 19504 × 10¹

Who releases notable AI models

In 2026 so far, 101 of 102 notable models came from companies alone. In 2015 it was 13 of 28.

Data: Epoch AI, CC BY 4.0. 2026 is a partial year.

Show the data as a table
YearIndustryBothAcademiaOtherUnited StatesMixed countriesChinaElsewhereClosed (API or app only)Open weightsUnreleased
200820402004010
200960507004002
201011804006003
201103804106004
201232803307006
20137210011206002
201457210124215015
201513690166060511
20162110180281011011118
2017261118017144201124
2018231040225191920
2019302412240190903137
202025169026115812821
2021462391332314994129
202250231715317615134136
2023693018266221912245936
20248214305372712443916
2025102521624341156478
202610110059134762365

What capability costs

68 models in the index have both a confirmed price and a score on Epoch AI's Capabilities Index, which combines dozens of benchmarks into one number that is comparable across generations.

The line joins the models that no other model beats on both price and score. Every model below it costs more for the same capability.

gpt-6.1-terra scores 167.7 for $2 per million tokens. gpt-6.2-astra, the top scorer at 174.2, costs $16: 8 times as much.

Pick a provider to see where its models sit.

Scores: Epoch AI, CC BY 4.0.

Show the data as a table
ModelProviderBlended price ($/M tokens)Epoch Capabilities IndexOn the line
gpt-6.2-astraOpenAI$16174.2Yes
Claude Fable 5.2Anthropic$16173.5
Gemini 4 CarbonGoogle$4169.8Yes
gpt-6-astraOpenAI$20169.2
Claude Opus 5.5Anthropic$8168.7
gpt-6.1-solOpenAI$4168.6
gpt-6.1-terraOpenAI$2167.7Yes
Claude Fable 5.1Anthropic$20167.1
Claude Sonnet 5.5Anthropic$4166.6
Gemini 4 ArgonGoogle$4166.6
Claude Opus 5Anthropic$10164.7
Claude Fable 5Anthropic$20164.2
gpt-6-solOpenAI$4163.8
gpt-5.6-solOpenAI$8163.1
gpt-5.5-proOpenAI$67.5162.8
Kimi K3.1 ProMoonshot$4161.4
gpt-5.6-terraOpenAI$4.5160.8
gpt-5.4-proOpenAI$67.5160.3
Claude Haiku 5.5Anthropic$0.20160.3Yes
gpt-5.5OpenAI$11.25160.2
Claude Opus 4.8Anthropic$10159.9
DeepSeek FlashDeepSeek$0.53159.4
Gemini 3.8 FlashGoogle$1.5158.8
Kimi K3Moonshot$6158.6
gpt-5.4OpenAI$5.63157.9
Grok 4.6xAI$3157.8
Claude Opus 4.7Anthropic$10157.5
gpt-5.2-proOpenAI$57.75157.3
GLM-5.3Z.ai$2.15157.1
Claude Sonnet 5Anthropic$4156.9
gpt-5.6-lunaOpenAI$0.45156.7
Grok 4.7xAI$3156.4
gpt-6-lunaOpenAI$0.20156.4
Claude Opus 4.6Anthropic$10156.4
Gemini 3.1 Pro PreviewGoogle$4.5156.3
DeepSeek V4 ProDeepSeek$1.98156
Gemini 3.6 FlashGoogle$1.5155.2
Grok 4.5xAI$3155
gpt-5.2OpenAI$4.81154.5
GLM-5.3-FlashZ.ai$0.24153.8
GLM-5.2Z.ai$2.15153
Claude Sonnet 4.6Anthropic$6152.9
Gemini 3 Flash PreviewGoogle$1.13152.3
Claude Opus 4.5Anthropic$10151.9
Kimi K2.6Moonshot$1.71151.6
gpt-5-proOpenAI$41.25151.3
Kimi K2.7 CodeMoonshot$1.71151.2
gpt-5.1OpenAI$3.44151
GLM-5.1Z.ai$2.15150.7
gpt-5OpenAI$3.44150
Grok 4.3xAI$1.56149.9
gpt-5.4-miniOpenAI$1.69149.4
Claude Sonnet 4.5Anthropic$6147.6
o3-proOpenAI$35147.5
o3OpenAI$3.5147.1
GLM-5Z.ai$1.55146.5
gpt-5.4-nanoOpenAI$0.46146.3
gpt-5-miniOpenAI$0.69146
Gemini 3.5 Flash-LiteGoogle$0.85145.7
Gemini 3.1 Flash-LiteGoogle$0.56145
Gemini 2.5 ProGoogle$3.44144.9
GLM-4.7Z.ai$1143.4
Claude Haiku 4.5Anthropic$2142.8
Gemini 2.5 FlashGoogle$0.85140.8
gpt-5-nanoOpenAI$0.14139.1Yes
gpt-4.1OpenAI$3.5136.2
Gemini 2.5 Flash-LiteGoogle$0.17133.1
gpt-4oOpenAI$4.38127.1

How fast models write

Gemini 3.5 Flash-Lite writes 455 tokens a second, 13 times as fast as Kimi K3.1 Pro at 35.

Data: Artificial Analysis.

Show the data as a table
ModelProviderTokens per secondSetting measured
Gemini 3.5 Flash-LiteGoogle455
Claude Haiku 5.5Anthropic240max
DeepSeek FlashDeepSeek177V4.1 Flash 1017, max
Gemini 3.8 FlashGoogle137high
Claude Sonnet 5.5Anthropic135max with fallback
gpt-6-lunaOpenAI117max
Claude Opus 5.5Anthropic98max with fallback
Grok 4.7xAI94high
gpt-6.1-terraOpenAI91max
Gemini 4 ArgonGoogle83high
gpt-6.1-solOpenAI76max
DeepSeek V4 ProDeepSeek740813, max
GLM-5.3Z.ai74max
gpt-6-astraOpenAI73max
Claude Fable 5.1Anthropic64xhigh with fallback
Kimi K3.1 ProMoonshot35

Do people prefer the models that test best?

Mostly, but not in order. Among the top 20 models by Epoch's index, Gemini 4 Argon moves from 10th to 2nd when people vote, and gpt-6.2-astra drops from 1st to 17th.

Data: Epoch AI and Arena.

Show the data as a table
ModelEpoch rankArena rankEpoch scoreArena rating
gpt-6.2-astra117174.21468
Claude Fable 5.225173.51503
Gemini 4 Carbon31169.81527
gpt-6-astra413169.21475
Claude Opus 5.553168.71506
gpt-6.1-sol69168.61483
gpt-6.1-terra715167.71474
Claude Fable 5.186167.11503
Claude Sonnet 5.5916166.61473
Gemini 4 Argon102166.61520
Claude Opus 5117164.71490
Claude Fable 5124164.21504
gpt-6-sol1319163.81458
gpt-5.6-sol148163.11485
Kimi K3.1 Pro1510161.41482
gpt-5.6-terra1618160.81468
Claude Haiku 5.51720160.31421
gpt-5.51812160.21481
Claude Opus 4.81911159.91482
DeepSeek Flash2014159.41475

Scores across six independent leaderboards

The top 18 models by Epoch's index. Darker cells rank higher among every priced model that leaderboard has scored; blank cells were not tested.

Data: Epoch AI, Artificial Analysis, Vals AI and Arena.

Show the data as a table
ModelEpoch Capabilities IndexArtificial Analysis Intelligence IndexVals Index v2.2Terminal-Bench 4.0GPQA DiamondArena Text
gpt-6.2-astra174.26267.763.11468
Claude Fable 5.2173.56571.365.21503
Gemini 4 Carbon169.86070.563.11527
gpt-6-astra169.2576559.61475
Claude Opus 5.5168.75970.165.21506
gpt-6.1-sol168.65563.955.11483
gpt-6.1-terra167.75160.549.51474
Claude Fable 5.1167.15058.193.41503
Claude Sonnet 5.5166.65570.964.11473
Gemini 4 Argon166.65069.457.61520
Claude Opus 5164.753.593.41490
Claude Fable 5164.241.493.21504
gpt-6-sol163.844.41458
gpt-5.6-sol163.137.995.21485
gpt-5.5-pro162.8
Kimi K3.1 Pro161.44662.563.61482
gpt-5.6-terra160.822.790.91468
gpt-5.4-pro160.3

66 AI tools by adoption

43 of the 66 tools in the index are open source.

Show the data as a table
ToolCategoryAdoptionOpen sourceEntry price
CursorDev tooling98No$8/mo
LangChainAgent frameworks96YesFree
GitHub CopilotDev tooling95No$10/mo
OllamaInfrastructure95YesFree
LangGraphAgent frameworks94YesFree
Claude CodeDev tooling93No$20/mo
vLLMInfrastructure91YesFree
PineconeInfrastructure90No$20/mo
CrewAIAgent frameworks88YesFree
LangSmithEvals86No$39/seat
LM StudioInfrastructure86No$20/mo
LlamaIndexAgent frameworks85YesFree
pgvectorInfrastructure85YesFree
OpenAI Agents SDKAgent frameworks84YesFree
UnslothInfrastructure84YesFree
LangfuseEvals83YesFree
RedisInfrastructure83YesFree
AutoGenAgent frameworks82YesFree
GroqInfrastructure82NoUsage
QdrantInfrastructure81YesFree
WindsurfDev tooling80No$20/mo
MilvusInfrastructure80YesFree
WeaviateInfrastructure79YesFree
PromptfooEvals78YesFree
MongoDB Atlas VectorInfrastructure78NoUsage
ChromaInfrastructure77YesFree
BraintrustEvals76No$249/mo
LettaAgent frameworks76YesFree
W&B WeaveEvals75YesFree
Browser UseAgent frameworks75YesFree
RagasEvals74YesFree
Together AIInfrastructure74NoUsage
HeliconeEvals73YesFree
SupermavenDev tooling73No$10/mo
Semantic KernelAgent frameworks72YesFree
Fireworks AIInfrastructure72NoUsage
Roo CodeDev tooling72YesFree
PortkeyEvals72YesFree
Arize PhoenixEvals71NoFree
AgnoAgent frameworks71YesFree
ClineDev tooling70YesFree
DeepEvalEvals69YesFree
Pydantic AIAgent frameworks68YesFree
OpenHandsDev tooling68YesFree
TruLensEvals68YesFree
DSPyAgent frameworks67YesFree
AiderDev tooling66YesFree
LanceDBInfrastructure66YesFree
turbopufferInfrastructure65No$16/mo
HaystackAgent frameworks64YesFree
GiskardEvals64YesFree
Google ADKAgent frameworks63YesFree
Replit AgentDev tooling62No$20/mo
MastraAgent frameworks62YesFree
GalileoEvals61No$100/mo
Amazon Q DeveloperDev tooling60No$19/mo
UpTrainEvals60YesFree
smolagentsAgent frameworks59YesFree
ContinueDev tooling58YesFree
Gemini Code AssistDev tooling57NoNo longer sold
Comet OpikEvals56YesFree
TabnineDev tooling55NoEnterprise only
AmpDev tooling54No$20/mo
Sourcegraph CodyDev tooling52NoEnterprise only
ZedDev tooling50YesFree
JetBrains AI AssistantDev tooling48No$10/mo

What the tools cost to start

44 of 66 are free to use. 13 more have a paid plan for one developer under $50 a month.

Show the data as a table
Entry priceToolsCount
FreeCline, Aider, Continue, Zed, LangChain, LangGraph, CrewAI, LlamaIndex, OpenAI Agents SDK, AutoGen, Semantic Kernel, Pydantic AI, DSPy, Haystack, Google ADK, smolagents, Langfuse, Promptfoo, W&B Weave, Ragas, Helicone, Arize Phoenix, DeepEval, Comet Opik, pgvector, Redis, Qdrant, Milvus, Weaviate, Chroma, LanceDB, Roo Code, OpenHands, Letta, Browser Use, Agno, Mastra, Portkey, TruLens, Giskard, UpTrain, Ollama, vLLM, Unsloth44
Pay as you goGroq, MongoDB Atlas Vector, Together AI, Fireworks AI4
Under $20/moCursor, GitHub Copilot, Amazon Q Developer, JetBrains AI Assistant, turbopuffer, Supermaven6
$20 to $49Claude Code, Windsurf, Replit Agent, LangSmith, Pinecone, Amp, LM Studio7
$50 and upBraintrust, Galileo2
Enterprise onlyTabnine, Sourcegraph Cody2
No longer soldGemini Code Assist1

Methods, dates and every source behind these charts are on the sources page.

Get an email when the index changes

Price changes, new tools, acquisitions and shutdowns, sent when the data is updated.