Comparison of generative AI models
From Wikipedia, the free encyclopedia
This is a comparison of generative AI models, such as language models and text-to-image models, across various benchmarks and user-generated Elo ratings. Leading large language models are sometimes called frontier models.
Background
Independent evaluation is widely used to compare frontier models because developer-reported scores are hard to compare directly.
- Artificial Analysis (AA) runs the Artificial Analysis Intelligence Index v4.3.2 across 10 evaluations including AA-Briefcase, GDPval-AA, AutomationBench-AA and Terminal-Bench 4.0, plus price per 1M tokens and output speed in tokens/s.[1][2] Its pricing and index data have been cited in coverage reporting DeepSeek V4-Flash as by far the cheapest well-known model to run.[3]
- Arena.ai, formerly Chatbot Arena from University of California, Berkeley, ranks models by blind human votes into elo-style ratings.[4][5] It tripled to a $1.7 billion valuation in January 2026 after raising $150 million.[4] Earlier coverage described its spin-out from an academic project into a $600 million startup.[6]
- Epoch AI publishes the Epoch Capabilities Index (ECI), a composite across 50 plus benchmarks scaled so that Claude 3.5 Sonnet equals 130.[7] GPT-6 Astra set a new record at 166 in September 2026, ahead of Claude Fable 5.1 at 164.[8] Its independence has been scrutinized over delayed disclosure of OpenAI funding for FrontierMath.[9]
- Vals AI publishes the Vals Index, agentic performance across finance, coding and legal tasks weighted by share of United States GDP, updated 10 September 2026 with 56 models tested.[10]
Language models
Text-to-image models
See also
Notes
- An asterisk denotes an Artificial Analysis estimated score pending independent evaluation; n/a indicates no published score in the cited snapshot.
- Scores shown are for each model's highest-performing reasoning effort level on Artificial Analysis; lower-effort configurations of the same model are omitted.
- Arena.ai scores shown are each model's best across reported effort variants on the leaderboard.
- Arena Elo ratings annotated as preliminary are based on fewer than 5,000 votes at the time of collection and are subject to greater statistical uncertainty as additional votes are recorded.