US agency CAISI: DeepSeek V4 Pro trails leading AI by 8 months
DeepSeek's latest model ranks behind GPT-5 but beats all other Chinese AIs.
The US government's CAISI (Center for AI Standardization and Innovation) evaluated DeepSeek V4 Pro, finding it eight months behind the most advanced US AI models. Released in April 2026 with 1.6 trillion parameters, it scored comparably to OpenAI's GPT-5 from August 2025. However, it achieved 200 points higher than Kimi K2.5 (China's previous best), representing a 3x task-solving probability improvement. CAISI also noted DeepSeek V4 Pro as more cost-effective than other models of comparable performance, outperforming GPT-5.4 mini in 5 of 7 benchmarks by 41–53%.
- DeepSeek V4 Pro benchmarks equivalent to GPT-5 from August 2025, lagging US leaders by eight months.
- Model scores 200 points above Kimi K2.5 on CAISI's scale (3x task-solving probability improvement).
- CAISI rates DeepSeek V4 Pro as the most cost-effective Chinese model, with US GPT-5.4 mini still top overall.
Why It Matters
Highlights the persistent US lead in AI while showing China's cost-effective, competitive alternative for enterprises.