DeepSeek V4 Pro GA Release Tested: Coding, 3D Modeling, and Game Generation Benchmarked

A Major Open-Source Model That Launched Without Fanfare

The general availability (GA) release of DeepSeek V4 Pro, codenamed "0813," has quietly gone live. Much like the earlier V4 Flash GA release, this launch came with almost no official fanfare — there's no announcement post on X, only a few references on the official website hinting that using the V4 Pro model identifier now routes you to the new 0813 version. The model has also appeared on OpenRouter, which effectively confirms this is a legitimate public release.

According to hands-on analysis by Bilibili creator Bijan Bowen, the most noteworthy aspect of this release is the magnitude of the capability jump. Both the V4 Pro preview and V4 Flash preview had felt like works-in-progress — which made sense given the "preview" label. But what made people excited was this: Flash made a "massive" leap from its preview to the 0731 GA release. If V4 Pro could follow the same trajectory, that would be something genuinely worth watching.

DeepSeek V4 Pro: Scale and Pricing

Based on available information, V4 Pro 0813 features a 1 million token context window with a maximum output length of 384K — a very impressive number. The model supports an optional thinking mode with adjustable reasoning intensity, though all coding tasks in this evaluation were run at maximum intensity to test peak performance.

If the model architecture matches the preview version, this is a 1.6 trillion parameter Mixture-of-Experts (MoE) model with 49 billion active parameters. While still enormous, it's notably leaner than models approaching 3 trillion parameters like Kimi K3 or Qwen 3.8 Max — a meaningful distinction for an open-weight model.

Pricing is arguably its biggest selling point: currently $0.0043 per million input tokens and $0.87 per million output tokens — almost absurdly cheap. DeepSeek has repeatedly hinted that API prices will be "significantly raised" soon, but even at 5–10x the current rate, it would still be highly competitive among comparable models.

According to what appears to be DeepSeek WeChat group benchmark data (unconfirmed officially), V4 Pro's DeepSWE score jumped from 12.8 on the preview to 62.7 on 0813 — if accurate, that would put it in a very strong position among open-source models.

What is MoE? Mixture-of-Experts (MoE) is a model architecture that splits the network into multiple "expert" sub-networks, activating only a small fraction during each inference pass. With 1.6 trillion total parameters but only 49 billion active, roughly 3% of parameters are used per forward pass — dramatically reducing per-inference compute costs. This allows MoE models to maintain massive "knowledge capacity" while keeping inference costs close to those of a dense model of equivalent active size. DeepSeek, Mixtral, Qwen, and other leading open-source models all use this architecture. It's precisely this sparse activation that allows V4 Pro to maintain extremely low API pricing despite its enormous parameter count.

Coding and Game Generation: Hands-On Tests

Bijan ran a series of zero-shot coding tests via Codex and the newly released ChatGPT desktop Beta (both routed to DeepSeek at maximum reasoning intensity).

Browser OS Simulation

The model generated a simulated desktop environment with a transparent background and synthwave aesthetic, complete with a file manager, notepad, calculator, settings, and games. The overall look was reminiscent of Qwen 3.8, with a built-in "holographic desktop" effect — described as looking "like an app demo inside an Apple Vision Pro headset." The file system is static, but support for creating and deleting files and neofetch display added nice touches.

Skate Game and Subway FPS

The C++ skateboarding game "Skate NYC" took about 68 minutes to complete. Compared to the "dismal" results from the preview version, this was described as a "significant improvement" — though issues like tilted building faces and skewed text remained, and the creator joked that "the trees look a bit like poop." Still, it was in a playable state overall.

For the subway FPS test, the model produced "the best subway car I've seen in testing" — clean and minimal, with smooth weapon models and even brass shell casings on the floor. The downside was the absence of enemies (zombies), and repeated self-debugging loops ultimately led to a disappointing final verdict.

On zero-shot testing: A "zero-shot" test means providing no examples or mid-task corrections — just a description, expecting the model to deliver in one pass. This is a rigorous standard for evaluating real-world generalization and engineering reliability, closer to a cold-start scenario in actual development. DeepSWE is a benchmark specifically designed to assess large models' software engineering capabilities, scoring them on autonomous bug fixes and feature development in real GitHub repositories. A jump from 12.8 to 62.7 — if confirmed — would mean the model's autonomous task completion rate on real engineering work (not simple code completion) rose from roughly 1 in 8 to more than 3 in 5: a qualitative leap.

3D Modeling and Frontend Design

RB26 Inline-6 Engine 3D Model

This was a high-difficulty task: generate a 3D-printable RB26 engine model requiring minimal supports and designed to house an N20 micro DC motor. The model spent about 1 hour and 10 minutes on it, conducting detailed research into component orientations (exhaust manifold below, intake on top, etc.) and even downloading images and specs for the motor — despite having no vision capability, which was puzzling.

The result was called "the best one I've seen in testing so far," with impressive assembly rendering — but the oil pan orientation issues made it unprintable. The creator's observation was astute: "The story isn't the model itself; it's the research it did behind the scenes."

Luxury Watch Frontend Website

The watch website test took 1 hour and 30 minutes, landing "very close to excellent." A wood-grain background, leather display mat, cinematic panning of 3D watch assets, and a dual-watch comparison view (stainless steel vs. gold, both rotatable) were all well-executed. The numeric hour markers even featured multi-material detail (gold/silver with black fill). The only letdown was the leather mat obscuring the view of the watch straps, but the overall premium feel was satisfying.

Cinematic Narrative Game Generation

The cinematic game "Steve the PC Repairman," completed via Codex in about 65 minutes, exceeded expectations. The game leveraged the OpenAI API to generate character voice assets, with character heads naturally swaying side to side during multi-character dialogue, and a soundtrack that fit each scene well. Though the puzzle pieces didn't fully assemble into the correct image, all the individual elements were there.

The final test, "Street Yeet," was an outright success — strong cartoon aesthetic, snappy pacing, fun physics interactions (launched objects collide with other scene elements), a combo multiplier, and a scoring system. Highly playable.

Summary: Outstanding Value with Room to Improve

The total cost for the entire test suite was just $1.91 — compared to roughly $32 for a similar Qwen 3.8 Max evaluation a few weeks prior. Even at 5x the current price, that's hard to call expensive.

Overall, V4 Pro 0813 is a "significant improvement" over the preview, shedding the "work-in-progress" feel of earlier releases. That said, the creator admitted the Flash model surprised him more on a performance-per-dollar basis (being much smaller), and V4 Pro faces stiffer competition at its tier.

The main complaint was the model being "too token-hungry" — at maximum reasoning intensity, it exhaustively re-checks every line of code, and in some cases attempted to launch headless browsers and take screenshots for heatmaps despite having no vision capability. This dramatically inflated both time and cost. For zero-shot testing, generating a result and iterating on human feedback is often more efficient. This may be a side effect of maximum reasoning intensity — something worth factoring into real-world use.

All things considered, as an extremely affordable open-weight model, DeepSeek V4 Pro 0813's transition from preview to GA is a meaningful step that the open-source community should be watching.

On "token hunger" and extended thinking: This issue is especially pronounced in high-intensity reasoning (extended thinking) mode. When the model is allowed to output large volumes of "chain-of-thought" tokens, it tends to exhaustively verify every assumption — including attempting to call tools like screenshots and heatmaps it has no actual ability to use. These attempts are guaranteed to fail but still consume significant token budget. Since reasoning tokens are typically billed as output tokens, extremely high thinking depth can multiply API costs dramatically. Users can set a thinking budget parameter (capping reasoning token output) to balance quality and cost rather than defaulting to maximum intensity — especially important when optimizing for value.