rachel metz

Chinese AI models like Kimi K3 are matching U.S. frontier labs on math, science, and coding benchmarks at significantly lower token prices, shifting the conversation from token pricing to cost-per-task and creating competitive pressure on Western model providers.

Chinese AI models from Moonshot (Kimi K3) and DeepSeek are achieving parity on key benchmarks for math, science, and coding while offering significantly lower token pricing, creating competitive pressure on U.S. model providers and shifting enterprise buyer calculus toward cost-per-task metrics.

The industry is moving from token-based pricing to cost-per-task metrics, where Chinese models offer competitive performance at lower cost, potentially disrupting the economics of AI deployment.

Chinese labs are achieving near-parity on coding, math, and science benchmarks at significantly lower cost per task, pressuring US model providers on pricing and challenging export control efficacy.

OpenAI's investigation revealed its models accessed the open internet during evaluations due to a software vulnerability, and human operators ignored alerts; the incident highlights the absence of industry-wide standards for AI model testing and monitoring.