Bijan Bowen tests the general-availability DeepSeek V4 Pro at maximum reasoning effort across a browser operating system, C++ skateboarding game, subway shooter, printable engine model, 3D watch site and two more ambitious games. The model has a very large context window and selectable reasoning, but its published benchmark picture was still uncertain when the review was recorded.
The weaker results exposed recurring reliability problems. The skate game was visually broken, the subway shooter failed to produce usable enemies after feedback and the detailed engine model was not practical to print without support. DeepSeek also spent substantial time trying to verify visual work without vision, making manual inspection and targeted feedback more efficient than letting it pursue elaborate self-checks.
Other tasks were much stronger. The model researched the engine's real component layout carefully, built a polished interactive watch presentation and produced a playable arcade-style city game. A cinematic game contained overlapping audio and geometric issues, but still showed unusually rich scene construction and character behaviour.
Bijan Bowen concludes that V4 Pro is a significant upgrade from the preview and inexpensive at the tested API price, though the smaller Flash variant had impressed him more relative to its size. The practical tradeoff is strong capability and low current cost against long runs, high token use and uneven zero-shot reliability.
Watch the original on YouTube