SemiAnalysis: Kimi K3 Ranks Third Globally, Could Reveal Hidden Profit Margins of OpenAI and Anthropic

SemiAnalysis analysts Jordan Nanos and Max Kan recently analyzed Kimi K3, the model developed by Chinese AI startup Moonshot AI, concluding that it outperforms Google Gemini in comprehensive benchmark tests. This not only reflects the narrowing gap between Chinese and U.S. AI models but also offers new insights into the business models of closed-source AI firms like Anthropic and OpenAI. According to SemiAnalysis’s overall assessment, Kimi K3 currently ranks third globally, trailing only Fable 5 and GPT-5.6, and surpassing Google Gemini. The analysts noted that while this result does not signal major issues for Google’s AI business, Kimi K3’s publicly disclosed parameter count, performance, and pricing provide a reference for external estimates of the economic value of closed-source models. Kimi K3 has 2.8 trillion parameters, far exceeding most open-source models. Jordan Nanos stated that a model of this size cannot be deployed on a single NVIDIA B200 GPU, requiring higher-spec hardware such as GB300, B300-class systems, or AMD MI355X. Based on this, he speculated that Anthropic and OpenAI’s flagship closed-source models likely operate at a similar parameter scale, rather than holding an order-of-magnitude advantage. In terms of business models, Kimi K3’s launch price is close to Anthropic’s Sonnet series: input pricing is approximately $3 per million tokens, and output pricing is around $15 per million tokens, a roughly threefold increase over the previous Kimi generation. Max Kan argued that if Moonshot AI is not operating at a long-term loss, then Anthropic and OpenAI charging higher prices for models of comparable size suggests their API business may have high profit margins. “Selling API tokens could be more profitable than SaaS,” he said. However, the two analysts emphasized that these judgments are not based on the AI companies’ public financial data, but rather on reverse inference drawn from Kimi K3’s parameters, pricing, and performance.