China's Kimi Nearly Catches the Frontier, Then Gets Caught in a Hype Storm

Moonshot AI, the Beijing startup behind the Kimi chatbot, just dropped Kimi K3.1 Pro, and on paper it's a big leap. The model has about 2.8 trillion parameters, handles text and images, and can take in a million tokens of context at once. On Artificial Analysis's Intelligence Index, a widely watched leaderboard, it jumped from 44 to 55 points, good enough for fourth place in the world, behind only Claude Opus 5.5, Claude Sonnet 5.5 and Meta's top model.

The coding numbers are what turned heads. On one tough terminal-based test, the previous Kimi scored around 13%. The new one scores in the mid-60s, almost level with the best American models. And it's cheap. Pricing is $2 per million input tokens and $10 per million output, a third lower than before and about half what Opus 5.5 costs. Moonshot says it will release the open weights on November 4, making it one of the strongest models anyone can download and run themselves. The company also says it deliberately stripped hacking-related data out of training for that public version, because it doesn't think open models should push the edge of cyberattack ability.

So why isn't the conversation all good news? An X user posted screenshots of what he says were instructions for a paid influencer campaign, asking creators to promote Kimi as a cheap alternative to Claude and, in one line, not to add the paid partnership label. He says he turned it down. The screenshots haven't been verified, and it's unclear whether Moonshot, an agency, or someone else wrote them. As of Saturday, Moonshot hadn't responded.

Even so, it's a fair moment to ask how much of the AI buzz in your feed is actually paid for.