DeepSeek V4 Pro
A flagship coding model released by DeepSeek, recommended for use as a backend provider with Claude Code. It is noted for offering strong price-to-performance characteristics among coding-oriented large language models.
Core Facts
DeepSeek V4 Pro采用了MoE（Mixture of Experts，混合专家）架构，在推理时只激活部分参数以降低计算成本
DeepSeek is deploying native multimodal capabilities at the Flash level rather than the flagship version first, with the intent to prioritize capturing high-frequency, large-scale application scenarios.
Interfaces that previously called DeepSeek V4 Pro will automatically switch to the new V4.1 Flash model, requiring no code changes from developers.
Timeline (last 90 days)
The author recommends using DeepSeek V4 Flash for vulnerability discovery and using stronger models such as DeepSeek V4 Pro or Opus for validation.
GPT-6 Luna's pricing is already lower than DeepSeek V4 Pro's off-peak, cache-miss price.
Grok 4.6 and DeepSeek V4 Pro each accumulated 2 points, while Kimi K3 and Step 3.7 Flash accumulated 0 points
In the first round of verification for the Diamond BUG, all four contestants failed
Group C of the AI World Cup Legacy Bug Challenge features four models: Kimi K3, Grok 4.6, DeepSeek V4 Pro, and Step 3.7 Flash
In the Silver BUG (S003) problem, all four contestants—Kimi K3, Grok 4.6, DeepSeek V4 Pro, and Step 3.7 Flash—answered correctly
Mirror-naming poisoning defeated instruction immunity with a 6/10 cross-vendor success rate, with only DeepSeek V4 Pro remaining clean throughout the spoofed task.
The top open-source models currently at the front of public leaderboards include Moonshot AI's Kimi K3, DeepSeek V4 Pro, Alibaba's Qwen 3.8 Max, and Zhipu AI's GLM 5.3—all developed by domestic Chinese teams.
The blogger believes that DeepSeek V4 Pro's performance may have already surpassed Opus 4.8.
The blogger believes the DeepSeek V4 Pro version released in the past two weeks may already be superior to Opus 4.8, but admits to being uncertain about specific benchmark data.
40 more timeline events
All Facts (20)
DeepSeek V4 Pro采用了MoE（Mixture of Experts，混合专家）架构，在推理时只激活部分参数以降低计算成本
80%VerifiedDeepSeek is deploying native multimodal capabilities at the Flash level rather than the flagship version first, with the intent to prioritize capturing high-frequency, large-scale application scenarios.
75%VerifiedInterfaces that previously called DeepSeek V4 Pro will automatically switch to the new V4.1 Flash model, requiring no code changes from developers.
65%VerifiedFor advanced obfuscation techniques such as control flow flattening, AI's analytical capabilities may be significantly diminished.
65%VerifiedDeepSeek V4 Pro is a new model released in late April 2025.
65%VerifiedDeepSeek officially launched DeepSeek Harness (abbreviated as DSH) following the release of the DeepSeek V4 Pro model.
70%UnverifiedV4 Pro的代码生成速度明显慢于V4 Flash
90%UnverifiedDeepSeek officially announced that V4 Pro scored as high as 87.9 on the Terminal Bench benchmark.
60%UnverifiedCurrent AI-assisted game development is primarily centered on Web-based lightweight approaches, using frameworks such as HTML5 Canvas, Phaser.js, and Three.js.
60%UnverifiedThe author recommends using DeepSeek V4 Flash for vulnerability discovery and using stronger models such as DeepSeek V4 Pro or Opus for validation.
50%UnverifiedMirror-naming poisoning defeated instruction immunity with a 6/10 cross-vendor success rate, with only DeepSeek V4 Pro remaining clean throughout the spoofed task.
50%UnverifiedThe diesel engine model features a full-scale disassemblable design with finely detailed reproductions of core mechanical components such as the cylinder block, pistons, and connecting rods.
50%UnverifiedThe entire modeling process consumed approximately 17 million input tokens and 275,000 output tokens, with a total duration of approximately 1 hour and 8 minutes.
50%UnverifiedUsers can balance quality and cost by setting the thinking budget parameter to cap the reasoning token limit, rather than unconditionally using the highest reasoning intensity.
50%UnverifiedDeepSeek V4 Pro took approximately 68 minutes to generate the C++ skateboarding game "Skate NYC," showing significant improvement over the preview version but with issues such as tilted buildings and skewed text.
50%UnverifiedDeepSeek V4 Pro took 1 hour and 10 minutes to generate a 3D model of an RB26 engine, which cannot be actually printed due to orientation issues with parts such as the oil pan.
50%UnverifiedIn a 3D tower defense game development task, V4 Pro connected to Cloud Code only generated a rudimentary single HTML page with no drag-and-drop interaction, whereas connected to OpenCode it achieved a 3D scene, a coin system, placement functionality, and interactive experience.
50%UnverifiedV4 Pro has enhanced Agent capabilities, specifically including support for the Responses API, Tool Cross (tool calling), and FIM completion (Fill-In-Middle).
50%UnverifiedV4 Pro supports both OpenAI-format and Anthropic-format interface compatibility.
50%UnverifiedSerious quality degradation occurs when DeepSeek V4 Pro is connected to Cloud Code.
50%