DeepSeek V4, China's Top Model

CAISI evaluated the open-weight model DeepSeek V4 Pro in April 2026 and estimated it to be about 8 months behind the US frontier.

The evaluation focused on 5 domains and 9 benchmarks, and the IRT-based comparison used 16 benchmarks and 35 models. DeepSeek V4 was classified as the best-performing model from the PRC that CAISI has evaluated so far.

• Its IRT-estimated Elo was 800 ±28, lower than GPT-5.5 (1260 ±28) and Anthropic Opus 4.6 (999 ±27).
• On the developer's own benchmarks, it appeared comparable to Opus 4.6 and GPT-5.4, but on CAISI's non-public benchmarks it was closer to GPT-5 level.
• It notably underperformed compared to US models on ARC-AGI-2 semi-private, PortBench, and CTF-Archive-Diamond in particular.
• In terms of cost, it was cheaper on 5 of 7 benchmarks compared to GPT-5.4 mini, ranging from 53% cheaper to 41% more expensive.
• CAISI served the model using H200/B200 GPUs and the developer's recommended settings, and reproduced the developer's self-reported results on GPQA-Diamond.