🔥 DeepSeek V4 Official Leaked: Outperforms Claude Opus 4

DeepSeek V4 official hasn't been formally announced yet, but the API has already silently switched to the official version via gray release. Multiple independent testers have run deep evaluations with a stunning conclusion: V4 outperforms Claude Opus 4, generating complete, playable 3D games from a single text prompt.

🚨 Key Finding
Based on independent tester feedback, DeepSeek V4 official's overall capability surpasses Claude Opus 4, placing it between Opus 4 and Gemini 2.5 Pro. This makes DeepSeek the third provider globally to achieve "super-Opus" capability — at 1/50th the price of Opus.

📊 V4 Official vs Global Top Models

ModelOverallCode GenPrice (Output/M tokens)Status
Claude Opus 4⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐$75Released
DeepSeek V4 Official⭐⭐⭐⭐☆⭐⭐⭐⭐⭐~$0.83 (off-peak $0.42)Gray release
Gemini 2.5 Pro⭐⭐⭐⭐☆⭐⭐⭐⭐$10Released
GPT-4o⭐⭐⭐⭐⭐⭐⭐⭐$10Released
Grok 4⭐⭐⭐⭐⭐⭐⭐⭐$15Released

Bottom line: DeepSeek V4 outperforms Opus 4 at 1/90th the price (off-peak) to 1/45th (peak). This is currently the world's highest value frontier model.

🎮 Real-World Evidence: One Prompt = Complete Game

Here are 10 independent Bilibili creator tests using DeepSeek V4 official (gray release), generating complete, runnable games from a single text prompt:

① Terraria / Binding of Isaac / RTS / Cut the Rope / Stardew Valley

Multi-genre generalization test covering 6 different game types. V4 performed stunningly.

🔗 Watch Test →

② Three Game Examples from Single Prompts

Three different game styles from single prompts, testing consistency and reliability.

🔗 Watch Test →

③ Backrooms & Horror Game from One Prompt

Backrooms-style horror game with atmosphere, jump scares, and level design.

🔗 Watch Test →

④ Minecraft Clone 🔥

The most impactful test: a complete 3D voxel game with block placement, destruction, and inventory systems. Core evidence for the "outperforms Opus 4" claim.

🔗 Watch Test →

⑤ CS:GO-Style FPS

FPS game from one prompt, including shooting, movement, and map design.

🔗 Watch Test →

⑥ Comparison: Gemini 3.5 Pro / Grok 4.5

Head-to-head comparison of V4 vs Gemini and Grok in code generation tasks.

🔗 Watch Test →

⑦ Super Mario

Classic platformer with jumping, movement, enemies, and coins.

🔗 Watch Test →

⑧ Super Mario 3D Racer

3D racing game showcasing V4's 3D rendering capabilities.

🔗 Watch Test →

⑨ Early Gray-Release Minecraft

Earlier gray-release version already produced impressive results. Official version improves further.

🔗 Watch Test →

⑩ One-Sentence Game Generation

Complete game from a single sentence with stunning results.

🔗 Watch Test →

💰 Super-Opus Capability at 1/90th the Price

BillingV4 Pro (Off-Peak)V4 Pro (Peak)V4 Flash (Off-Peak)V4 Flash (Peak)Claude Opus 4
Cache Hit¥0.025¥0.05¥0.02¥0.04$3.75
Cache Miss¥3¥6¥1¥2$15
Output¥6¥12¥2¥4$75

Per million tokens. DeepSeek in RMB, Claude in USD.

🤔 What This Means

  • New era of AI coding: from "code assistant" to "one-sentence product builder"
  • Chinese AI surpasses Opus for the first time: DeepSeek becomes the 3rd provider with super-Opus capability
  • Price disruptor: super-Opus at 1/90th the price — developer's dream
  • Official launch imminent: expected mid-July 2026
⚠️ Note
Tests based on gray-release version; official release may include further optimizations. Capability assessments are based on independent tester comparisons, not official benchmark data. DeepSeek V4 official expected to fully launch mid-July 2026.