A U.S. government institute, the Center for AI Standards and Innovation (CAISI), has concluded that China's most powerful AI, DeepSeek V4 Pro, lags behind the frontier by about eight months. CAISI does not average benchmark scores but uses Item Response Theory to estimate each model's latent capability across nine benchmarks in five domains. The open-weight flagship model scored around 800 (±28), which is very close to GPT-5.4 mini. However, the analysis revealed that the gap is actually decreasing. The AI developer under the pseudonym Ex0bit countered that there was no "gap" and that the best model’s score was the reference point to see how capable a model is.