Local LLM coding agents benchmarked against Claude Opus and Codex

The author tested various local AI models against cloud models like Anthropic's Claude Opus and OpenAI's Codex using a standardized banking code task. The tests revealed that while large local models running on high-end hardware like the NVIDIA DGX Spark can approach cloud-level performance, smaller models suitable for laptops fall significantly short. Meanwhile, numerous Chinese cloud models through OpenRouter delivered competitive performance at a fraction of the cost. Open article →