Artificials
Compare models
AI TRENDS

How AI is changing.

The frontier of capability, what it costs, where models are made and how fast new ones arrive, from published results and the model catalog.

At a glance

Where AI stands now, from published results and the model catalog.

Frontier

Claude Opus 5.5 holds the highest capability index (167.3), with OpenAI and Moonshot AI close behind.

Open weights

The best open-weights model, Kimi K3 (157.4), is 9.9 points behind the best closed model, Claude Opus 5.5.

Countries

Models from the United States lead at 167.3; the best from China reaches 157.4, 9.9 points behind.

Cost

Reaching index 150 or better cost $3.4 per 1M tokens with GPT-5; DeepSeek V4 Flash 0731 does it for $0.17, 20 times less, at today’s list prices.

Releases

342 notable models were released in the last twelve months, about 29 a month.

The frontier

The highest capability index each maker has reached, and when.

Highest capability index to date, by maker

Each step is a model that beat the maker’s own best. Select a maker in the legend to hide or show it.

Capability index. Use the arrow keys to move between points.

Highest capability index to date, by maker
SeriesDateModelCapability index
AnthropicJul 11, 2023Claude 2120.1
AnthropicAug 9, 2023Claude Instant120.3
AnthropicFeb 29, 2024Claude 3 Opus126.9
AnthropicJun 20, 2024Claude 3.5 Sonnet130.0
AnthropicOct 22, 2024Claude 3.5 Sonnet (October 2024)133.6
AnthropicFeb 24, 2025Claude 3.7 Sonnet141.2
AnthropicMay 22, 2025Claude Opus 4142.7
AnthropicAug 5, 2025Claude Opus 4.1144.1
AnthropicSep 29, 2025Claude Sonnet 4.5146.8
AnthropicNov 24, 2025Claude Opus 4.5150.1
AnthropicFeb 5, 2026Claude Opus 4.6155.2
AnthropicApr 16, 2026Claude Opus 4.7156.3
AnthropicMay 28, 2026Claude Opus 4.8158.2
AnthropicJun 9, 2026Claude Fable 5162.1
AnthropicJul 24, 2026Claude Opus 5162.8
AnthropicSep 1, 2026Claude Fable 5.1164.7
AnthropicSep 22, 2026Claude Opus 5.5167.3
OpenAIMar 14, 2023GPT-4 (Mar 2023)125.9
OpenAIJan 25, 2024GPT-4 Turbo (Nov 2023)126.5
OpenAIApr 9, 2024GPT-4 Turbo (Apr 2024)127.3
OpenAIMay 13, 2024GPT-4o (May 2024)129.0
OpenAISep 12, 2024o1 Mini135.8
OpenAIDec 17, 2024o1141.9
OpenAIApr 16, 2025o3146.9
OpenAIJun 10, 2025o3 Pro147.4
OpenAIAug 7, 2025GPT-5150.0
OpenAIOct 7, 2025GPT-5 Pro150.3
OpenAIDec 11, 2025GPT-5.2 Pro155.4
OpenAIFeb 5, 2026GPT-5.3 Codex156.8
OpenAIMar 5, 2026GPT-5.4 Pro158.9
OpenAIApr 23, 2026GPT-5.5 Pro162.1
OpenAISep 3, 2026GPT-6 Astra166.4
Moonshot AIJul 12, 2025Kimi K2 (Jul 2025)140.1
Moonshot AINov 6, 2025Kimi K2 Thinking146.0
Moonshot AIJan 27, 2026Kimi K2.5148.0
Moonshot AIApr 20, 2026Kimi K2.6151.1
Moonshot AIJul 16, 2026Kimi K3157.4
GoogleDec 13, 2023Gemini 1.0 Pro117.0
GoogleMay 14, 2024Gemini 1.5 Pro (May 2024)126.9
GoogleSep 24, 2024Gemini 1.5 Pro (Sept 2024)131.7
GoogleDec 11, 2024Gemini 2.0 Flash (Dec 2024)134.7
GoogleJan 21, 2025Gemini 2.0 Flash Thinking (Jan 2025)135.4
GoogleMar 31, 2025Gemini 2.5 Pro (Mar 2025)144.2
GoogleJun 5, 2025Gemini 2.5 Pro (Jun 2025)145.3
GoogleNov 18, 2025Gemini 3 Pro152.9
GoogleFeb 19, 2026Gemini 3.1 Pro154.8
GoogleAug 13, 2026Gemini 3.7 Flash157.3
Meta AIFeb 24, 2023Llama 65B110.2
Meta AIJul 18, 2023Llama 2-70B113.8
Meta AIApr 18, 2024Llama 3-70B122.9
Meta AIJul 23, 2024Llama 3.1-405B128.8
Meta AIApr 5, 2025Llama 4 Scout129.6
Meta AIApr 6, 2025Llama 4 Maverick132.2
Meta AIApr 8, 2026Muse Spark152.0
Meta AIJul 9, 2026Muse Spark 1.1154.2
Meta AIAug 5, 2026Muse Spark 1.2154.9
Meta AISep 2, 2026Muse Spark 1.3156.8
xAIDec 12, 2024Grok-2 (Dec 2024)130.5
xAIApr 9, 2025Grok 3138.3
xAIJun 24, 2025Grok-3 mini140.3
xAIJul 9, 2025Grok 4146.4
xAIFeb 17, 2026Grok 4.20152.0
xAIJul 8, 2026Grok 4.5153.9
xAIAug 12, 2026Grok 4.6156.4
AlibabaSep 24, 2023Qwen 14B113.0
AlibabaJun 7, 2024Qwen2 72B125.3
AlibabaSep 17, 2024Qwen2.5 32B128.5
AlibabaSep 19, 2024Qwen2.5 72B129.0
AlibabaJan 25, 2025Qwen2.5 Max132.5
AlibabaMar 5, 2025QwQ 32B137.6
AlibabaApr 28, 2025Qwen3 235B A22B139.3
AlibabaJul 25, 2025Qwen3-235B-A22B-Thinking (Jul 2025)143.8
AlibabaFeb 13, 2026Qwen3.5 397B-A17B146.7
AlibabaFeb 16, 2026Qwen 3.5 Plus (hosted 397B-A17B)146.8
AlibabaMar 31, 2026Qwen 3.6 Plus147.7
AlibabaApr 20, 2026Qwen 3.6 Max (Preview)149.2
AlibabaMay 19, 2026Qwen3.7 Max153.7
AlibabaAug 2, 2026Qwen 3.8 Max156.4
Z.aiDec 22, 2025GLM-4.7143.5
Z.aiFeb 11, 2026GLM-5145.8
Z.aiApr 7, 2026GLM-5.1149.8
Z.aiJun 16, 2026GLM-5.2151.8
Z.aiAug 14, 2026GLM-5.3155.6

Epoch Capabilities Index, published by Epoch AI (CC BY 4.0), by release date. Makers with at least three models on the index.

Each maker’s best model

Highest capability index per maker.
  1. Claude Opus 5.5167.3
  2. GPT-6 Astra166.4
  3. Kimi K3157.4
  4. Gemini 3.7 Flash157.3
  5. Muse Spark 1.3156.8
  6. Grok 4.6156.4
  7. Qwen 3.8 Max156.4
  8. GLM-5.3155.6
  9. DeepSeek V4 Pro 0813155.3
  10. MiniMax M3146.9
  11. Mistral Medium 3.5141.3
  12. Phi-4130.4

Open weights and closed models

How close downloadable models come to the best closed ones.

Highest capability index to date: open weights and closed

Open-weights models can be downloaded and run anywhere; closed models are only offered by their makers and hosts.

Capability index. Use the arrow keys to move between points.

Highest capability index to date, open weights and closed
SeriesDateModelCapability index
Closed modelsMar 14, 2023GPT-4 (Mar 2023)125.9
Closed modelsJan 25, 2024GPT-4 Turbo (Nov 2023)126.5
Closed modelsFeb 29, 2024Claude 3 Opus126.9
Closed modelsApr 9, 2024GPT-4 Turbo (Apr 2024)127.3
Closed modelsMay 13, 2024GPT-4o (May 2024)129.0
Closed modelsJun 20, 2024Claude 3.5 Sonnet130.0
Closed modelsSep 12, 2024o1 Mini135.8
Closed modelsDec 17, 2024o1141.9
Closed modelsMar 31, 2025Gemini 2.5 Pro (Mar 2025)144.2
Closed modelsApr 16, 2025o3146.9
Closed modelsJun 10, 2025o3 Pro147.4
Closed modelsAug 7, 2025GPT-5150.0
Closed modelsOct 7, 2025GPT-5 Pro150.3
Closed modelsNov 18, 2025Gemini 3 Pro152.9
Closed modelsDec 11, 2025GPT-5.2 Pro155.4
Closed modelsFeb 5, 2026GPT-5.3 Codex156.8
Closed modelsMar 5, 2026GPT-5.4 Pro158.9
Closed modelsApr 23, 2026GPT-5.5 Pro162.1
Closed modelsJul 24, 2026Claude Opus 5162.8
Closed modelsSep 1, 2026Claude Fable 5.1164.7
Closed modelsSep 3, 2026GPT-6 Astra166.4
Closed modelsSep 22, 2026Claude Opus 5.5167.3
Open-weights modelsFeb 24, 2023Llama 65B110.2
Open-weights modelsJul 18, 2023Llama 2-70B113.8
Open-weights modelsJul 20, 2023Stable Beluga 2117.1
Open-weights modelsNov 2, 2023Yi 34B117.4
Open-weights modelsDec 11, 2023Mixtral 8x7B118.5
Open-weights modelsApr 17, 2024Mixtral 8x22B122.0
Open-weights modelsApr 18, 2024Llama 3-70B122.9
Open-weights modelsMay 7, 2024DeepSeek-V2 (MoE-236B, May 2024)124.8
Open-weights modelsJun 7, 2024Qwen2 72B125.3
Open-weights modelsJul 23, 2024Llama 3.1-405B128.8
Open-weights modelsSep 19, 2024Qwen2.5 72B129.0
Open-weights modelsDec 12, 2024Phi-4130.4
Open-weights modelsDec 26, 2024DeepSeek V3132.3
Open-weights modelsJan 20, 2025DeepSeek R1139.0
Open-weights modelsApr 28, 2025Qwen3 235B A22B139.3
Open-weights modelsMay 28, 2025DeepSeek-R1 (May 2025)141.3
Open-weights modelsJul 25, 2025Qwen3-235B-A22B-Thinking (Jul 2025)143.8
Open-weights modelsSep 29, 2025DeepSeek V3.2 Exp145.0
Open-weights modelsNov 6, 2025Kimi K2 Thinking146.0
Open-weights modelsDec 1, 2025DeepSeek V3.2146.3
Open-weights modelsJan 27, 2026Kimi K2.5148.0
Open-weights modelsApr 7, 2026GLM-5.1149.8
Open-weights modelsApr 20, 2026Kimi K2.6151.1
Open-weights modelsJun 16, 2026GLM-5.2151.8
Open-weights modelsJul 16, 2026Kimi K3157.4

Weight access as recorded by Epoch AI. Data: Epoch AI (CC BY 4.0).

Countries

Where the most capable models are made, by the developer’s country.

Highest capability index to date, by country

Countries with at least five models on the index.

Capability index. Use the arrow keys to move between points.

Highest capability index to date, by country
SeriesDateModelCapability index
United StatesFeb 24, 2023Llama 65B110.2
United StatesMar 14, 2023GPT-4 (Mar 2023)125.9
United StatesJan 25, 2024GPT-4 Turbo (Nov 2023)126.5
United StatesFeb 29, 2024Claude 3 Opus126.9
United StatesApr 9, 2024GPT-4 Turbo (Apr 2024)127.3
United StatesMay 13, 2024GPT-4o (May 2024)129.0
United StatesJun 20, 2024Claude 3.5 Sonnet130.0
United StatesSep 12, 2024o1 Mini135.8
United StatesDec 17, 2024o1141.9
United StatesMar 31, 2025Gemini 2.5 Pro (Mar 2025)144.2
United StatesApr 16, 2025o3146.9
United StatesJun 10, 2025o3 Pro147.4
United StatesAug 7, 2025GPT-5150.0
United StatesOct 7, 2025GPT-5 Pro150.3
United StatesNov 18, 2025Gemini 3 Pro152.9
United StatesDec 11, 2025GPT-5.2 Pro155.4
United StatesFeb 5, 2026GPT-5.3 Codex156.8
United StatesMar 5, 2026GPT-5.4 Pro158.9
United StatesApr 23, 2026GPT-5.5 Pro162.1
United StatesJul 24, 2026Claude Opus 5162.8
United StatesSep 1, 2026Claude Fable 5.1164.7
United StatesSep 3, 2026GPT-6 Astra166.4
United StatesSep 22, 2026Claude Opus 5.5167.3
ChinaJun 1, 2023Baichuan1 7B90.5
ChinaSep 6, 2023Baichuan2 13B103.2
ChinaSep 24, 2023Qwen 14B113.0
ChinaNov 2, 2023Yi 34B117.4
ChinaMay 7, 2024DeepSeek-V2 (MoE-236B, May 2024)124.8
ChinaJun 7, 2024Qwen2 72B125.3
ChinaSep 17, 2024Qwen2.5 32B128.5
ChinaSep 19, 2024Qwen2.5 72B129.0
ChinaDec 26, 2024DeepSeek V3132.3
ChinaJan 20, 2025DeepSeek R1139.0
ChinaApr 28, 2025Qwen3 235B A22B139.3
ChinaMay 28, 2025DeepSeek-R1 (May 2025)141.3
ChinaJul 25, 2025Qwen3-235B-A22B-Thinking (Jul 2025)143.8
ChinaSep 29, 2025DeepSeek V3.2 Exp145.0
ChinaNov 6, 2025Kimi K2 Thinking146.0
ChinaDec 1, 2025DeepSeek V3.2146.3
ChinaJan 27, 2026Kimi K2.5148.0
ChinaApr 7, 2026GLM-5.1149.8
ChinaApr 20, 2026Kimi K2.6151.1
ChinaMay 19, 2026Qwen3.7 Max153.7
ChinaJul 16, 2026Kimi K3157.4
FranceSep 27, 2023Mistral 7B v0.1112.2
FranceDec 11, 2023Mixtral 8x7B118.5
FranceFeb 26, 2024Mistral Large122.0
FranceApr 17, 2024Mixtral 8x22B122.0
FranceJul 24, 2024Mistral Large 2 (Jul 2024)127.5
FranceNov 18, 2024Mistral Large 2 (Nov 2024)128.5
FranceMay 7, 2025Mistral Medium 3134.1
FranceApr 28, 2026Mistral Medium 3.5141.3

Developer countries from Epoch AI’s model records. Data: Epoch AI (CC BY 4.0).

The cost of capability

What it costs to use a model at least as capable as each level.

Cheapest list price to reach each capability level

Each step is a new model that reaches the level for less. Log scale; lower is cheaper.

USD per 1M tokens. Use the arrow keys to move between points.

Cheapest list price to reach each capability level
SeriesDateModelUSD per 1M tokens
Index 120 or higherJan 25, 2024GPT-4 Turbo (Nov 2023)$15
Index 120 or higherFeb 26, 2024Mistral Large$6.5
Index 120 or higherApr 17, 2024Mixtral 8x22B$3
Index 120 or higherApr 18, 2024Llama 3-70B$0.57
Index 120 or higherJul 18, 2024GPT-4o mini$0.26
Index 120 or higherDec 12, 2024Phi-4$0.22
Index 120 or higherJan 20, 2025DeepSeek R1 Distill Qwen 14B$0.15
Index 120 or higherMar 12, 2025Gemma 3 12B$0.075
Index 120 or higherJul 27, 2026Qwen3.7 Flash$0.055
Index 130 or higherSep 12, 2024o1 Mini$1.9
Index 130 or higherDec 12, 2024Phi-4$0.22
Index 130 or higherJan 20, 2025DeepSeek R1 Distill Qwen 14B$0.15
Index 130 or higherAug 5, 2025GPT OSS 20B$0.087
Index 130 or higherJul 27, 2026Qwen3.7 Flash$0.055
Index 140 or higherDec 17, 2024o1$26.3
Index 140 or higherJan 31, 2025o3 Mini$1.9
Index 140 or higherMay 20, 2025Gemini 2.5 Flash (May 2025)$0.89
Index 140 or higherJun 17, 2025Gemini 2.5 Flash (Jun 2025)$0.85
Index 140 or higherJun 24, 2025Grok-3 mini$0.35
Index 140 or higherSep 29, 2025DeepSeek V3.2 Exp$0.3
Index 140 or higherDec 1, 2025DeepSeek V3.2$0.26
Index 140 or higherFeb 25, 2026Qwen 3.5 Flash (hosted 35B-A3B)$0.17
Index 140 or higherApr 2, 2026Gemma 4 26B A4B$0.11
Index 140 or higherJun 4, 2026Nemotron 3 Ultra$0.1
Index 140 or higherJul 27, 2026Qwen3.7 Flash$0.055
Index 150 or higherAug 7, 2025GPT-5$3.4
Index 150 or higherDec 17, 2025Gemini 3 Flash$1.1
Index 150 or higherJul 9, 2026GPT-5.6 Luna$0.45
Index 150 or higherJul 31, 2026DeepSeek V4 Flash 0731$0.17

Prices are today’s list prices from models.dev (MIT), blended three input tokens to one output token, not the price each model launched at. Capability: Epoch Capabilities Index.

Context windows

How much text new models take in at once.

Median context window of new models, by quarter

Notable releases from known makers. Log scale.

Tokens. Use the arrow keys to move between points.

Median context window of new models, by quarter
SeriesDateModelTokens
Closed modelsNov 15, 20232023 Q416.4K
Closed modelsFeb 15, 20242024 Q1128K
Closed modelsMay 15, 20242024 Q2128K
Closed modelsAug 15, 20242024 Q3131.1K
Closed modelsNov 15, 20242024 Q4200K
Closed modelsFeb 15, 20252025 Q1198.3K
Closed modelsMay 15, 20252025 Q2200K
Closed modelsAug 15, 20252025 Q3256K
Closed modelsNov 15, 20252025 Q4262K
Closed modelsFeb 15, 20262026 Q1262.1K
Closed modelsMay 15, 20262026 Q2262.1K
Closed modelsAug 15, 20262026 Q31M
Closed modelsNov 15, 20262026 Q4327.7K
Open-weights modelsFeb 15, 20242024 Q1131.1K
Open-weights modelsMay 15, 20242024 Q2129.5K
Open-weights modelsAug 15, 20242024 Q3128K
Open-weights modelsNov 15, 20242024 Q4128K
Open-weights modelsFeb 15, 20252025 Q1131.1K
Open-weights modelsMay 15, 20252025 Q2129.5K
Open-weights modelsAug 15, 20252025 Q3131.1K
Open-weights modelsNov 15, 20252025 Q4204.8K
Open-weights modelsFeb 15, 20262026 Q1204.8K
Open-weights modelsMay 15, 20262026 Q2262.1K
Open-weights modelsAug 15, 20262026 Q31M

Context windows as listed in the models.dev catalog (MIT), by release date. Quarters need at least three models in a group.

Training scale

How much computation went into training notable models.

Estimated training compute, by release date

Floating-point operations, on a log scale. The line marks the largest run so far.

  • Alibaba
  • EleutherAI
  • DeepSeek
  • Meta
  • Microsoft
  • OpenAI
  • Google
  • NVIDIA
  • Z.ai
  • Moonshot AI
  • xAI
  • Anthropic
  • Mistral AI
  • Xiaomi
  • Amazon
  • MiniMax
  • Largest so far

Use the arrow keys to move between models. Press Enter to open the selected model. A table with the same data follows the chart.

Training compute is Epoch AI’s estimate (CC BY 4.0); for closed models it is inferred from public information and can be uncertain.

The pace of releases

How many notable models arrive each month, and who ships the most.

Notable releases per month

Models from known makers that the maker sells, several hosts list, or a test covers.

Notable model releases per month
Nov 202416
Dec 202427
Jan 202530
Feb 20259
Mar 202517
Apr 202537
May 202519
Jun 202518
Jul 202529
Aug 202521
Sep 202532
Oct 202527
Nov 202516
Dec 202531
Jan 202621
Feb 202642
Mar 202646
Apr 202645
May 202620
Jun 202626
Jul 202627
Aug 202623
Sep 202641
Oct 20264

Release dates as listed in the models.dev catalog (MIT).

Most releases in the last twelve months

Notable models per maker.
  1. OpenAI54
  2. Alibaba48
  3. Google43
  4. Z.ai24
  5. Anthropic23
  6. xAI23
  7. ByteDance21
  8. DeepSeek19
  9. Mistral AI19
  10. Xiaomi14
Sources and method

Every chart is calculated by Artificials from published data: records are the models that beat every earlier model in their group, by release date.