Open-weight AI models remain 5–6 months behind closed-source leaders like Claude Code; Chinese labs face resource constraints while U.S. companies stabilize the open ecosystem.

The gap between open and closed AI models is measured not by benchmarks alone but by real-world utility in agentic workflows. Claude Code reached a performance milestone in December 2025; open models have not yet achieved equivalent capability at comparable price points like $5/month, and the author expects this gap to persist closer to 12+ months.

American AI companies—including Nvidia with Nemotron, Google with Gemma, and others—are stabilizing the open model ecosystem with permissive licenses (e.g., Gemma 4 adopted Apache 2.0 License). Gemma 4's models are tying or outperforming equivalently sized Qwen 3.5/3.6 models, which had been the default open model at those sizes.

Chinese labs including Kimi, Z.ai, DeepSeek, and Qwen are heavily resource limited and lack an immediate path to scaling training processes like large U.S. labs. In contrast, Epoch AI data shows U.S. compute distribution: Google 25%, Meta 11%, OpenAI 11%, Anthropic 6%.

Open models are expected to serve automated enterprise agents and low-cost domains rather than drive modern knowledge work, while Claude Code and Codex remain the current best path to massive AI revenue growth.