China’s DeepSeek Upgrades V4 Pro: Claude Fable Is Only 5% Better at 4,500% the Price
DeepSeek quietly released the finalized version of its flagship model, updating the API name to DeepSeek-V4-Pro-0813 without a public announcement. The pricing stayed the same: about $0.435 per million input tokens and $0.87 per million output tokens. Earlier results and benchmarks had been based on a preview build, not the finished model. DeepSeek shared its own comparisons across 10 agent benchmarks, where gaps versus leading models are generally small. The company claims the main tradeoff is cost: V4 Pro is vastly cheaper than premium U.S. models, with blended rates around $0.65 per million tokens versus roughly $30 for Fable 5. That makes DeepSeek about 46 times cheaper on token pricing, and in some per-task estimates over 100 times cheaper. The catch is that the new model has not yet been independently benchmarked. Some tests are internal and not publicly verifiable, but the model weights are MIT-licensed and available on Hugging Face, so outside evaluation is possible.
