China’s latest AI closes gap with US models (again)

Kimmy from Full House 🤝 Kimi from China 🤝 barging right in. A Beijing-based artificial intelligence lab called Moonshot surprised Silicon Valley yesterday when it released Kimi K3, a new AI model that it says beats most US systems and costs less to use.According to benchmarks provided by Moonshot:Kimi outperforms all available AI models except for two—Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol—both of which came out in the last two months.Kimi can store far more information than leading US models, boasting 2.8 trillion parameters (which are like an AI model’s brain cells) compared with an estimated 1.5 trillion for Claude Opus 4.8, released in May.Price comparison: Moonshot will charge businesses $15 per million output tokens (a metric of AI usage), compared to costs of $30 for GPT-5.6 Sol and $50 for Claude Fable 5.Threat Level Midnight: Kimi is also supposedly better at most coding and agentic tasks than other recently released versions of GPT and Claude, while still being cheaper. That’s especially threatening to OpenAI and Anthropic, because their models’ coding capabilities are major revenue drivers, and both companies are eyeing IPOs. Kimi’s birth announcement rattled markets in Taiwan, Japan, and the US, where the tech-heavy Nasdaq recorded its biggest dip of the week.Déjà vu…China’s DeepSeek similarly spooked Silicon Valley last year, when it released open-source AI that it claimed went toe-to-toe with US models and undercut their prices. Moonshot also plans to open-source Kimi later this month, meaning developers will be able to download and customize it.Chinese President Xi Jinping applauded his country’s progress on low-cost AI yesterday and called for “a symphony of global cooperation,” panning the US for its efforts to restrict China’s access to AI technology.But US companies claim China isn’t playing fair: Anthropic recently accused Chinese tech giant Alibaba (one of Moonshot’s investors) of attempting to extract Claude’s AI capabilities to train its own models, and OpenAI has made similar claims against DeepSeek.—ML