We break down the new Chinese AI model that’s already winning over coders and surpassing Anthropic’s flagship.
Chinese startup Moonshot AI has released the open-source Kimi K3 model with 2.8 trillion parameters, beating Anthropic’s flagship Claude Fable 5 in the Arena frontend development ranking. Kimi K3 scored 1,679 points against Fable 5’s 1,631, taking first place in six of seven categories–from marketing pages and dashboards to consumer apps.
Hot topic: Bitcoin Bottom May Not Arrive Until Q4, Analyst Says
The model is already available on the website, Kimi Work, Code, and API. Full weights and a report will be released on July 27. Kimi K3 costs $3 per 1 million input tokens and $15 for output tokens–compared to Fable 5’s $10 and $50.
Contents
Why Kimi K3 Beat Claude Fable 5 and GPT-5.6 Sol
Kimi K3’s leap surprised the industry. Its predecessor, Kimi K2.6, ranked only 18th on the Arena leaderboard. The new model jumped to first place, beating Fable 5 by 48 points. By comparison, Fable 5 beat third-place GPT-5.6 by just 13 points.
Anthropic’s only win over Kimi K3 came in the “games” category. Still, Anthropic maintains a strong presence in the top 20, with nine models on the list.
Kimi K3 is built on a new Kimi Delta Attention and Attention Residuals architecture, which processes long sequences more efficiently and selectively extracts information from layers. The model also uses Stable LatentMoE, where only 16 of 896 experts are active at once–delivering a 2.5x efficiency gain over the previous version.
Read more: Sentient (SENT) Price Prediction — Can This AI Crypto Hit a New All-Time High in 2026?
Price and Availability: Why Kimi K3 Is Reshaping the AI Market
The real story is price. Kimi K3 costs $3 per 1 million input tokens and $15 for output. Fable 5 charges $10 and $50 respectively. That makes the Chinese model far more accessible for developers and enterprise clients. Open-sourcing the weights by July 27 will let programmers run frontier AI on their own hardware for free.
Kimi K3 can handle long engineering sessions with almost no human input, navigate large repositories, and manage terminal tools. In a GPU kernel optimization test, the model worked independently for up to 24 hours across four tasks, matching Fable 5’s performance and beating Opus 4.8 and GPT-5.6 Sol. In one case, K3 reproduced the universal I-Love-Q relations from computational astrophysics in two hours–compared to one to two weeks for a researcher.
Read more: What is AI in Crypto? Complete Guide to AI Agents, Tokens & Use Cases
Market Reaction and Competition With US Giants
Kimi K3’s release comes as competition intensifies between Chinese and US AI developers. Alibaba ordered employees to stop using Claude Code on July 10 over security concerns. Still, Anthropic remains influential, with nine models in the Arena top 20.
The events confirm a global trend: Chinese AI products are starting to outperform US competitors in monthly generation volume and accessibility. Open models from China are becoming a real alternative to proprietary systems–especially for developers who care about cost.
Learn more: China Crypto Regulation — Why Bitcoin Is Effectively Banned in the World’s Largest Market
