DeepSeek V4 Pro official release ships with 1M context, 384K output, beats Fable 5 on multiple agent benchmarks

DeepSeek released the official V4-Pro-0813 model at midnight Aug 13 with 1M context and 384K output, supporting JSON Output, Tool Calls, Responses API and Anthropic API. Benchmark gains vs the preview are large, including DeepSWE 12.8 to 62.7 and DSBench-Hard 31.1 to 67.2, surpassing Opus 4.8 and roughly tied with Fable 5 on multiple agent evaluations. Pricing remains 3x of V4-Flash, with concurrency limited to 500.