The Open Source AI Model Beating GPT-5 on Agentic Performance

Quick Overview

The open-source AI model Kimi K2 Thinking is outperforming proprietary models like GPT-5 and Claude Sonnet in agentic tasks, demonstrating a significant shift in AI development where Chinese open-source models are gaining ground on US counterparts by offering superior performance at a fraction of the cost, as evidenced by its performance on benchmarks and its ability to run locally.

Key Points: Kimi K2 Thinking scored 91% on the Humanity's Last Exam, surpassing GPT-5 and every other model mentioned, including DeepSeek V3, at a cost of $0.14/$0.29 per million tokens. The open-source lag is now measured in months, not years, with DeepSeek R1 following in 4 months, and GPT-5's lead over Kimi K2 shrinking to just 3 months. The 'closed model advantage window' has collapsed from 18+ months to 3-4 months, indicating a rapid catch-up by open-source models. The cost advantage is significant: Kimi K2 costs $0.16/$2.50 per million tokens compared to GPT-5's $1.25/$10.00 per million tokens for similar performance levels. The model can generate a full novel from one prompt, running up to 300 sequential tool calls per session, demonstrating advanced agentic capabilities. The release of Kimi K2 Thinking and subsequent models like MiniMax M2 suggests that the era of 'bigger models' is over, replaced by an era of 'smarter inference' and cost-effective local deployment. The video highlights that Chinese models are now competitive, with MiniMax's M2 ranking high on OpenRouter and Cognition AI's coding agent based on a Chinese model.

Context: The video discusses the rapidly evolving landscape of AI model development, focusing heavily on the competitive advancements made by Chinese open-source AI models challenging established US leaders like OpenAI and Anthropic. Key figures like Jensen Huang (Nvidia CEO) and various AI researchers/developers on X (formerly Twitter) are referenced to illustrate the closing performance and cost gap, particularly in agentic capabilities and coding benchmarks.