SemiAnalysis: Kimi K3 Ranks Third Globally, Unveiling OpenAI and Anthropic's Hidden Profit Potential

BlockBeats News, July 20th, Analysts Jordan Nanos and Max Kan from SemiAnalysis recently conducted an analysis of the Kimi K3 model under the Dark Side of the Moon, suggesting that the model outperformed Google Gemini in a comprehensive benchmark test. This not only reflects the narrowing AI model gap between China and the U.S., but also provides a new perspective on the business models of closed-source AI companies such as Anthropic and OpenAI.

According to SemiAnalysis' comprehensive evaluation, Kimi K3 is currently ranked third globally, behind only Fable 5 and GPT-5.6, and ahead of Google Gemini. The analysts stated that while this result does not indicate a significant issue with Google's AI business, Kimi K3's publicly disclosed parameter scale, performance, and pricing provide a reference for estimating the economic value of closed-source models.

Kimi K3 has 28 trillion parameters, far exceeding most public models. Jordan Nanos suggested that a model of this scale cannot be deployed on a single Nvidia B200 GPU and would require higher-spec hardware support such as GB300, B300-level systems, or AMD MI355X. Based on this, he speculated that Anthropic and OpenAI's flagship closed-source models may also have similar parameter scales, rather than having a magnitude advantage.

In terms of the business model, Kimi K3's launch price is close to Anthropic's Sonnet series models, with an input price of around $3 per million tokens and an output price of around $15 per million tokens, representing a threefold increase from the previous generation Kimi model.

Max Kan believes that if the Dark Side of the Moon is not operating at a long-term loss, then the fact that Anthropic and OpenAI are charging higher prices for similar-scale models implies that their API business may have a higher profit margin. "Selling API tokens may be more profitable than SaaS," he said.

However, the two analysts emphasized that the above assessments are not based on AI companies' public financial data but are deduced through reverse engineering based on Kimi K3's parameters, prices, and performance.