The AI model market in charts: who leads, how fast open models are catching up, where the best models come from, and what they cost. From LMArena’s published ratings since Mar 24, every model on the board at the time, and each model’s provider data.

Open or proprietary

1 of the top 20 models on LMArena has open weights in Oct 26, against 9 in Mar 24.

Top 20 models on LMArena, by weights

Oct 26Open weights 1Proprietary 19All periods: Open weights 27% · Proprietary 73% of 640 places

The gap

The best open-weight model trails the best proprietary one by 35 rating points, against 105 at Mar 7, 2024.

Best LMArena rating: open weights against proprietary

LatestBest proprietary 1526Best open weights 1491

Where the best models come from

In Oct 26, 17 of the top 20 come from labs in the United States and 3 from labs in China.

Top 20 models on LMArena, by the country of their lab

Oct 26United States 17China 3Europe 0Other 0All periods: United States 66% · China 25% · Europe 3% · Other 6% of 640 places

Which labs lead

Google holds the most places in the top 20 in Oct 26, with 7.

Top 20 models on LMArena, by lab

Oct 26OpenAI 0Google 7Anthropic 6Alibaba 2xAI 0DeepSeek 0Meta 4Other labs 1All periods: OpenAI 20% · Google 17% · Anthropic 16% · Alibaba 10% · xAI 6% · DeepSeek 6% · Meta 6% · Other labs 19% of 640 places

Releases

Models released per month

Oct 26Open weights 6Proprietary 6All periods: Open weights 51% · Proprietary 49% of 177 models

Prices

Blended price per million tokens, 187 models

$10+Open weights 0Proprietary 13All periods: Open weights 51% · Proprietary 49% of 187 models

Context windows

Context window, 187 models

1M+Open weights 20Proprietary 54All periods: Open weights 51% · Proprietary 49% of 187 models

How these are made

What is “the top 20”?

At each month’s last LMArena update, the 20 models with the highest text rating, counting every model on the board at the time, including those since retired. A model rated at several effort levels counts once, at its best.

Why do the charts start in Mar 24?

LMArena’s published history leaves the proprietary API models out of its earliest snapshots (OpenAI’s from before December 2023, Anthropic’s from before March 2024), which would make open models look ahead. The charts start at Mar 7, 2024, the first update with OpenAI, Anthropic and Google all on the board.

How is a model’s country decided?

By where its lab is headquartered. Models whose lab we can’t place count as Other.

Are these usage numbers?

No. These charts measure quality ratings, releases and prices. Usage on Nexus will be published once there is enough of it to report without exposing any one customer.