DeepSeek V4 Pro coding benchmark

How DeepSeek V4 Pro performed

DeepSeek V4 Pro by DeepSeek took the PomBench AI coding benchmark: every model gets the same prompt and writes self-contained browser code for a twelve-second scene: a Pomeranian runs across a backyard, jumps a puddle, enters its doghouse, retrieves a bone and turns toward the viewer. The video above is an independent offline browser recording of the code DeepSeek V4 Pro returned, captured at 960 × 540 and 30 fps, with no organizer edits to the code.

In this DeepSeek V4 Pro coding test the model produced 5,244 output tokens, including provider-reported reasoning (rank 32 of 46 by output tokens), and a verified API price is unavailable for this run, so no cost is shown. Is DeepSeek V4 Pro good at coding? One task cannot answer that on its own, so this LLM coding benchmark lets you watch the result instead of reading a score: check the gait, the puddle jump and reflection, the doorway occlusion and whether the bone stays in the dog's mouth.

Then compare DeepSeek V4 Pro vs DeepSeek V4.1 Flash and GPT-5.5 below, or open the AI model comparison to play up to four models on one synchronized timeline.