🔥 Gray Release Hands-On: A Suspected DeepSeek V4 Official Builds Complete Games in One Prompt
Multiple users report that DeepSeek's API and web app appear to be gray-testing a higher-capability model. Because DeepSeek has not officially confirmed the exact version, we refer to it here as the “V4 official gray-release build.” Independent testers generated fully playable games from a single sentence — Minecraft, Binding of Isaac, Backrooms, CS:GO, across 2D platformers, 3D voxel, horror, strategy, racing and more. Below: screenshots, chat-share links, and full videos — all publicly verifiable.
Across several front-end game-generation cases, testers felt the gray-release build outperforms Claude Opus 4.8, landing somewhere between Opus 4.8 and Fable 5.0. Caveat: without a shared prompt set, a unified runtime, or an independent benchmark, this only reflects the “one-prompt web game” scenario and cannot be treated as an overall capability verdict. For pricing context, see DeepSeek V4's time-based surge pricing.
🔬 Evidence & Method: How Trustworthy Are These Claims?
This is not an official benchmark — it's a verifiable evidence roundup. So you can judge for yourself, here are the sources and the limits of the methodology:
| Evidence type | Count | Confidence | Notes |
|---|---|---|---|
| DeepSeek official web chat shares | 4 | Higher | Generated on chat.deepseek.com; full chain-of-thought visible, hard to fake |
| OpenCode API logs | 9 | Medium | Third-party agent calling the DeepSeek API; confirms the API flipped, but prompts/runtime not fully public |
| Third-party Bilibili videos | 12 | Medium | Real hands-on demos, but mostly edited; raw prompts and failure counts not shared |
| Independent reproduction by us | Not yet | — | We plan to publish fixed prompts, model ID, token usage and cost |
Known limits: the evidence above comes from different testers with inconsistent prompts, browsers, model IDs and rounds, and we can't confirm every case hit the same gray build. So our conclusions apply to “this gray build's front-end game-generation performance” and shouldn't be extrapolated to code, reasoning, or long-form tasks. For broader model cross-evals, see DeepSeek V4 real-world tests (vs Opus 4.8 / Gemini 2.5 Pro).
📸 Real Screenshots (High Impact)
The following captures come from the gray-release version. All generated from a single text prompt, no manual edits:
① Horror Game (Atmosphere Maxed)
Lighting, jump scares and level structure on par with commercial horror titles.
② Compound Bow (SVG Physics Animation)
Pure SVG bow physics: draw, tension, parabolic arrow, hit feedback — all from one prompt.
③ 3D Scene Capture 1
2560×1440 high-res 3D scene with detail, materials and lighting at pro level.
④ 3D Scene Capture 2
Another high-res 3D scene showcasing V4's stability across complex builds.
⑤ No Man's Sky × Minecraft: Pseudo-Seamless Landing
Deep space flight and voxel planet approach — super pseudo-seamless landing demonstrates V4's 3D open-world generation.
⑥ Minecraft × No Man's Sky + Interstellar Portal
Stargate-style portal in a voxel open world — cross-dimensional teleportation generated from one prompt.
💬 DeepSeek Chat Share Links (Official Evidence)
These are DeepSeek's official web chat-share links. Open them to view the full chain-of-thought and generation trace — primary evidence that this is not faked with Fable 5:
- � Interactive buoyancy lab — sliders adjust object/liquid density, real-time Canvas animation of floating/suspending/sinking with Archimedes' principle explained
� DeepSeek Share 1 → - � No Man's Sky in three.js — procedurally generated planets, space flight, seamless landing, atmospheric scattering, sci-fi HUD
� DeepSeek Share 2 → - 🪐 3D Planetarium in three.js — all 8 planets + Saturn's rings + asteroid belt, orbit & rotation animations, mouse interaction, info labels
🔗 DeepSeek Share 3 → - ✈️ 3D Pacific Naval Battle — first-person flight, carrier takeoff, air-to-air/air-to-ship missiles + cannon + bombs, AI enemy planes & ships
🔗 DeepSeek Share 4 →
🛠️ OpenCode Share Links (API Hands-On)
These are conversations run through OpenCode against the gray-release DeepSeek V4 API, confirming the API side has already been switched to the official version:
- � Japanese horror game (Chilla's Art style) — convenience store night shift, PS1/VHS retro aesthetic, first-person flashlight, ghost AI
�🔗 OpenCode Share 1 → - 🌌 Stellaris web edition — procedural galaxy, empire management, resource economy, colonization & research, fleet combat
🔗 OpenCode Share 2 → - 🔫 CS:GO web edition — 3D FPS, Dust2-style map, buy menu, bomb defusal, AI enemies, recoil & spray patterns
🔗 OpenCode Share 3 → - ⛏️ Minecraft Survival Edition — survival/creative modes, day/night cycle, zombies skeletons spiders, crafting system, mining & building
🔗 OpenCode Share 4 → - � 3D Backrooms horror — infinite procedural maze, yellow wallpaper & carpet, VHS effect, flashlight battery, sound-driven monster AI
�🔗 OpenCode Share 5 → - ⚔️ Risk of Rain 2 — 3D third-person roguelike, item stacking, monster director, teleporter stage progression
🔗 OpenCode Share 6 → - � Angry Birds — slingshot physics, destructible structures, multiple bird abilities, level progression, Web Audio SFX
�🔗 OpenCode Share 7 → - 🌌 Minecraft + Deep Space — voxel aesthetic meets No Man's Sky gameplay, procedural planets, ship landing animation, resource gathering
🔗 OpenCode Share 8 → - 🚀 Minecraft × No Man's Sky (Interstellar Portal) — voxel open world + cross-dimensional stargate
🔗 OpenCode Share 9 →
🎬 Bilibili Video Compilation (Full Demos)
These are Bilibili creator tests using DeepSeek V4 official gray release, each generating a complete playable game from one prompt:
① Terraria / Binding of Isaac / RTS / Cut the Rope / Stardew Valley
Multi-genre generalization test — V4 stunned across six game types.
② Three Game Examples from Single Prompts
Three different styles verifying consistency and reliability.
③ Backrooms & Horror Game from One Prompt
Atmosphere, jump scares, full level design.
④ Minecraft Clone 🔥
Complete 3D voxel game: block placement, destruction, inventory — the core "beats Opus 4.8" evidence.
⑤ CS:GO-Style FPS
FPS from one prompt: shooting, movement, map design.
⑥ Head-to-Head: Gemini 3.5 Pro / Grok 4.5
China's pride reborn — cross-eval against Gemini 3.5 Pro and Grok 4.5.
⑦ Suspected Official: Super Mario
Classic platformer with jumping, movement, enemies and coins.
⑧ Super Mario 3D Racer
3D racer showcasing V4's rendering capabilities.
⑨ Gray-Release Minecraft
Even early gray-release results were strong; official version pushes further.
⑩ One-Sentence Full Game
Gray-released, one sentence in, real results out.
⑪ No Man's Sky × Minecraft: Pseudo-Seamless Planet Landing
Deep space exploration with voxel planets — super pseudo-seamless landing showcases V4's 3D open-world generation.
⑫ Minecraft × No Man's Sky + Portal Teleportation
Interstellar portal teleportation in a voxel open world — seamless cross-dimensional travel generated from one prompt.
🧠 Deep Analysis: What It Means
- A stunning leap from the preview: vs the V4 preview, the jump is comparable to R1-0120 → R1-0528.
- Reportedly ahead of GLM 5.2: in game generation it holds its own against GPT and Claude (Fable 5 still leads), but overall capability still needs an independent benchmark — don't over-extrapolate.
- Wish-coding transformation: DeepSeek's weak spot — especially Flash on wish coding — is now fast and accurate when given a full plan.
- Web version equally impressive: expert model is back to Pro (after a brief, fast-but-weaker Flash period).
- Pricing confidence: DeepSeek V4's API pricing stays the best value among Chinese models. Read against the two-tier split of China's AI model market, DeepSeek is trying to own both the “cheap infrastructure” and “frontier capability” layers at once.
💾 Minecraft × No Man's Sky Open-Source Download
The "Minecraft × No Man's Sky" build generated via DeepSeek V4 has been open-sourced — download and run it yourself:
These tests are based on the gray-release version, not official benchmarks. Whether it's the “official version” or “an even higher tier” is still debated. Given AI's winner-takes-most dynamics, the market widely suspects this is the official version — though it still awaits DeepSeek's confirmation. Full launch expected mid-July 2026.