Intelligence.
In perspective.

• Model rankings

  Every model on every published test, ranked one test at a time.

  Gemini 4 Argon ( high )Google1524

  Highest rating on Arena text, where people vote between two anonymous answers.

• Coding agents

  Coding agents and the models inside them, on real programming work.

  GPT-5.5Run by NexAU-AHE84.7%

  Highest share of the 89 Terminal-Bench 2.0 tasks solved.

• Model finder

  Pick a task, a budget and what the model must do, and get the best match.

  The page has the latest results.

• Image models

  Making and editing images, ranked by the votes of people comparing them.

  GPT-Image-2.5 SunburstOpenAI1427

  Highest rating on Arena text to image, from people’s votes.

• Video models

  Text to video, image to video and video editing, ranked by votes.

  Gemini Omni 1.1 FlashGoogle1516

  Highest rating on Arena text to video, from people’s votes.

• AI trends

  How fast capability rises and prices fall, and who is catching up.

  GPT-6 LunaOpenAI$0.20

  Cheapest model scoring 150 or more on the index, per 1M tokens. GPT-5 costs 17 times as much.

Results from Arena (CC BY 4.0), Terminal-Bench 2.0 (Apache-2.0, compiled by Epoch AI) and Epoch AI (CC BY 4.0); prices from models.dev (MIT).

• Highest capability index

  

  Claude Opus 5.5Anthropic

  166.8

  Epoch Capabilities IndexTest result

• Best at coding

  

  Claude Sonnet 5.5 (max effort)Anthropic

  91.4%

  LiveBench codingTest result

• Best at agentic coding

  

  DeepSeek V4.1 Flash (max)DeepSeek

  77.3%

  LiveBench agentic codingTest result

• Best at advanced math

  

  GPT-6 Astra (max)OpenAI

  93.7%

  FrontierMath Tiers 1–3Test result

• Best at science reasoning

  

  Gemini 4 Argon (high)Google

  96.4%

  GPQA DiamondTest result

• Best at real-world code fixes

  

  Claude Opus 4.7 (max)Anthropic

  83.5%

  SWE-bench VerifiedTest result

• Best factual recall

  

  Gemini 4 Argon (high)Google

  76.0%

  SimpleQA VerifiedTest result

• Best image generation

  

  Nano Banana 2Google

  85.3%

  T2I-CoReBenchTest result

• Best open-weights model

  

  Kimi K3Moonshot AI

  157.4

  Epoch Capabilities IndexTest result

• Lowest price in the top 10

  

  GPT-6.1 SolOpenAI

  $4.00

  Blended list price per 1M tokensCalculated

• Largest context window

  

  Llama 4 Scout 17B InstructMeta

  10M tokens

  Listed context windowListing data

• Newest release

  

  DeepSeek V4.1 Flash HereticDeepSeek

  Oct 31, 2026

  Listed release dateListing data