A user found that DeepSeek V4 Pro, while significantly cheaper than Claude Sonnet 4, performs nearly as well in practical coding tasks. The user developed a custom harness, cwcode, to bridge the remaining performance gap, particularly in areas like long-horizon planning and handling less-than-ideal code, where Claude still holds an advantage. However, DeepSeek V4 Pro excels in executing precise specifications and handling numerical/scientific code, often outperforming Claude in these specific areas. AI
IMPACT Highlights cost-performance trade-offs and the impact of custom harnesses on LLM coding capabilities.
RANK_REASON User experience report on model performance and cost-effectiveness, including custom tooling.
Read on HN — claude cli stories →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →