Is the second best AI agent lab OpenAI or a Chinese one?

Moonshot released Kimi K3 in mid-July at the World AI Conference in Shanghai. Four days later, Arena listed it on the leaderboard ranking AI labs by how much their models improve real-world work. It scored higher than its own lab's previous best than any model has before or since. However, it has been sliding ever since. Its top tier peaked at 10.60% on August 13 and reads 8.74% today.

The slide is the only reason OpenAI is back in second place, and Polymarket's second-best AI agent lab market pays 88 cents for it to stay there until the table is read on September 30. We are betting against that at 13 cents.

OpenAI has shipped nothing since GPT 5.6 in July. Its best model was listed at 10.94% and reads 9.76%, so it has fallen too, just like Moonshot. The gap is 1.02 points. In the Labs view this market resolves on, both are marked rank 2 to 3.

So, logically, being head-to-head, second place is taken by the lab that ships next. Moonshot has released a new Kimi every 59 days at the median since July 2025 and has never gone longer than 83. Today is day 48.

Arena publishes its whole history, and it takes a median of six days to list a model after release, so anything Moonshot ships before about September 24 is on the board when the check happens. Although the outperformance is not guaranteed. Typically, a newly listed model beats its own lab's previous best 17 times against 21.

Still, there's another reason we think 'OpenAI' 'yes' contract holders won't be happy with. OpenAI's new major model family, Astra, which was named on August 6, will likely be announced at its developer conference on September 29, the day before this market is read. This means Arena won't have enough time to list it on its platform, nor will users have enough time to properly evaluate it.