概要
DeepSeek V4 Pro is a 1.6-trillion-parameter Mixture-of-Experts model (49B active per token) released on April 24, 2026, as part of a two-tier open-weight lineup alongside V4 Flash. It introduces a hybrid Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) architecture that dramatically improves long-context efficiency — requiring only 27% of V3.2's inference FLOPs and 10% of its KV cache at the full 1-million-token context window. Pre-trained on 32T+ tokens with a comprehensive post-training pipeline featuring domain-specific expert cultivation followed by unified consolidation, V4 Pro ships under the MIT license with weights on HuggingFace and ModelScope.
The model's strongest performance is in coding and competitive programming: it leads all public models on LiveCodeBench (93.5% Pass@1), Codeforces (3206 rating), and ties Claude Opus 4.6 and Gemini 3.1 Pro on SWE-Bench Verified (80.6%). On the Artificial Analysis Intelligence Index it scores 52, making it the #2 open-weight reasoning model behind only Kimi K2.6. Its GDPval-AA agentic score of 1554 leads all open-weight models, and its MCPAtlas Public score of 73.6 nearly matches Claude Opus 4.6's 73.8. However, it shows meaningful gaps in factual knowledge (HLE 37.7%, SimpleQA 57.9%) and high hallucination rates (94% on AA-Omniscience), reflecting training emphasis on code and math over broad world knowledge.
Pricing is the model's most disruptive feature. At $1.74/M input (cache miss) and $3.48/M output — roughly 7-8x cheaper than Claude Opus 4.7 and GPT-5.5 on output tokens — V4 Pro collapses the price floor for frontier-adjacent quality. The companion V4 Flash model ($0.14/$0.28) provides 85-95% of V4 Pro's quality at 12x lower cost for most workloads. Combined with automatic context caching and a 50% off-peak discount, DeepSeek has created the most cost-efficient frontier-adjacent model stack available. The trade-offs are real — higher latency, political censorship in weights, regulatory concerns, and knowledge gaps — but for teams building production coding agents and extraction pipelines, V4 Pro represents the strongest quality-per-dollar proposition in the market as of July 2026.
ベンチマーク＆性能
DeepSeek V4 Pro (in its Max reasoning effort mode) delivers competitive frontier performance across most benchmark categories, with clear leadership in coding and competitive programming. Below is a detailed comparison against the primary closed frontier models.
### Coding & Software Engineering
| Benchmark | V4-Pro Max | Claude Opus 4.6 | GPT-5.4 | Gemini 3.1 Pro |
|---|---|---|---|---|
| SWE-Bench Verified | 80.6% | 80.8% | 80.0% | 80.6% |
| LiveCodeBench Pass@1 | **93.5%** | 88.8% | — | 91.7% |
| Codeforces Rating | **3206** | — | 3168 | 3052 |
| SWE-Bench Pro | 55.4% | 57.3% | 57.7% | 54.2% |
| SWE Multilingual | 76.2% | 77.5% | — | — |
V4 Pro leads on competitive programming (LiveCodeBench, Codeforces) and ties on real-world SWE-Bench Verified. It trails slightly on SWE-Bench Pro and SWE Multilingual, where Claude Opus 4.6 holds a 2-3 point edge.
### Knowledge & Reasoning
| Benchmark | V4-Pro Max | Claude Opus 4.6 | GPT-5.4 | Gemini 3.1 Pro |
|---|---|---|---|---|
| MMLU-Pro | 87.5% | 89.1% | 87.5% | **91.0%** |
| GPQA Diamond | 90.1% | 91.3% | 93.0% | **94.3%** |
| HLE (Humanity's Last Exam) | 37.7% | 40.0% | 39.8% | **44.4%** |
| SimpleQA-Verified | 57.9% | 46.2% | 45.3% | **75.6%** |
| Chinese-SimpleQA | 84.4% | 76.4% | 76.8% | **85.9%** |
Knowledge is V4 Pro's weakest category. It trails Gemini 3.1 Pro by significant margins on GPQA Diamond (4.2 pts), HLE (6.7 pts), and SimpleQA (17.7 pts). It does outperform Claude and GPT-5.4 on SimpleQA, suggesting different knowledge profiles rather than uniform superiority.
### Mathematics
| Benchmark | V4-Pro Max | Claude Opus 4.6 | GPT-5.4 | Gemini 3.1 Pro |
|---|---|---|---|---|
| HMMT Feb 2026 | 95.2% | 96.2% | **97.7%** | 94.7% |
| IMOAnswerBench | 89.8% | 75.3% | **91.4%** | 81.0% |
| Apex | 38.3% | 34.5% | 54.1% | **60.9%** |
| Apex Shortlist | **90.2%** | 85.9% | 78.1% | 89.1% |
Math performance is strong but not uniformly leading. IMOAnswerBench at 89.8% is well ahead of Claude (75.3%) and Gemini (81.0%), though GPT-5.4 edges ahead at 91.4%. Apex Shortlist at 90.2% is the highest score. HMMT trails GPT-5.4 by 2.5 points.
### Agentic & Tool Use
| Benchmark | V4-Pro Max | Claude Opus 4.6 | GPT-5.4 | Gemini 3.1 Pro |
|---|---|---|---|---|
| Terminal-Bench 2.0 | 67.9% | 65.4% | **75.1%** | 68.5% |
| MCPAtlas Public | 73.6% | **73.8%** | 67.2% | 69.2% |
| BrowseComp | 83.4% | 83.7% | 82.7% | **85.9%** |
| Toolathlon | 51.8% | 47.2% | **54.6%** | 48.8% |
| GDPval-AA (Elo) | 1554 | 1619 | **1674** | 1314 |
V4 Pro ties Claude on MCPAtlas (73.6 vs 73.8%) and leads open-weight models on agentic tasks. However, GPT-5.4 dominates Terminal-Bench (75.1% vs 67.9%) and GDPval-AA (1674 vs 1554), indicating a real gap for autonomous long-running agent workflows.
### Long-Context
| Benchmark | V4-Pro Max | Claude Opus 4.6 | Gemini 3.1 Pro |
|---|---|---|---|
| MRCR 1M (MMR) | 83.5% | **92.9%** | 76.3% |
| CorpusQA 1M | 62.0% | **71.7%** | 53.8% |
Long-context retrieval is competitive but trails Claude Opus 4.6 significantly on both MRCR 1M (9.4 pts) and CorpusQA 1M (9.7 pts). V4 Pro substantially outperforms Gemini 3.1 Pro on both metrics. BenchLM ranks V4 Pro #14 overall (provisional score 78) and #7 on verified benchmarks. Its category breakdown: Coding #10/101 (91st percentile), Agentic #21/120 (83rd percentile), Knowledge #26/107 (76th percentile).
詳細比較
### DeepSeek V4 Pro vs Claude Opus 4.6
| Dimension | V4-Pro | Claude Opus 4.6 |
|---|---|---|
| Output Price | $3.48/1M | $25.00/1M |
| SWE-Bench Verified | 80.6% | 80.8% |
| LiveCodeBench | 93.5% | 88.8% |
| Terminal-Bench 2.0 | 67.9% | 65.4% |
| MRCR 1M | 83.5% | 92.9% |
| Latency (avg) | ~28s | ~5.8s |
| License | MIT (open weights) | Closed |
V4 Pro matches Claude on SWE-Bench Verified and leads on competitive programming, at roughly 1/7th the output cost. Claude holds significant advantages in latency (5x faster), long-context quality, factual knowledge, and multi-file refactoring consistency. For batch/async coding workloads, V4 Pro offers dramatically better economics. For latency-sensitive user-facing applications or tasks requiring high factual accuracy, Claude remains the safer choice.
### DeepSeek V4 Pro vs GPT-5.4/5.5
| Dimension | V4-Pro | GPT-5.4 |
|---|---|---|
| Output Price | $3.48/1M | $15.00/1M |
| Codeforces | 3206 | 3168 |
| GPQA Diamond | 90.1% | 93.0% |
| IMOAnswerBench | 89.8% | 91.4% |
| Terminal-Bench 2.0 | 67.9% | 75.1% |
| Context Window | 1M tokens | 1M tokens |
V4 Pro leads on competitive programming (Codeforces 3206 vs 3168) and is 4-5x cheaper on output. GPT-5.4 pulls ahead on Terminal-Bench (75.1% vs 67.9%), GPQA Diamond (93.0% vs 90.1%), and HMMT math (97.7% vs 95.2%). GPT-5.5 launched with 82.7% on Terminal-Bench, widening the agentic gap further. For autonomous agent workloads that spend extended time in terminal environments, the GPT-5.x family maintains a clear advantage.
### DeepSeek V4 Pro vs Gemini 3.1 Pro
| Dimension | V4-Pro | Gemini 3.1 Pro |
|---|---|---|
| Output Price | $3.48/1M | $12.00/1M |
| SWE-Bench Verified | 80.6% | 80.6% |
| LiveCodeBench | 93.5% | 91.7% |
| GPQA Diamond | 90.1% | 94.3% |
| HLE | 37.7% | 44.4% |
| SimpleQA-Verified | 57.9% | 75.6% |
| MRCR 1M | 83.5% | 76.3% |
These models are SWE-Bench Verified twins (80.6% each), but their strengths diverge sharply. Gemini dominates knowledge-heavy benchmarks (GPQA +4.2 pts, HLE +6.7 pts, SimpleQA +17.7 pts) while V4 Pro leads on competitive programming (LiveCodeBench +1.8 pts, Codeforces +154 rating) and long-context retrieval (MRCR 1M +7.2 pts). V4 Pro is roughly 3.5x cheaper on output. Choose Gemini for research-assistant and factual Q&A workloads; choose V4 Pro for coding, extraction, and cost-sensitive production pipelines.
コミュニティ評価
The developer and research community response to DeepSeek V4 Pro has been strongly positive on technical merit, with important caveats on trust and deployment.
**Coding community adoption:** Developers on Hacker News and Reddit report that V4 Pro 'holds the thread of what happened three hours ago' in extended agentic sessions, attributing this to the new CSA+HCA attention architecture. Multiple practitioners have switched production coding workloads from V3.2, reporting 'noticeably better multi-file refactors' though acknowledging a gap to Claude Opus 4.7 on complex cross-file consistency. One practitioner on DeepSeek AI Guide noted: 'For agentic coding at a one-seventh output cost, it is the rational pick this week.'
**Benchmark skepticism:** Several independent reviewers (Artificial Analysis, LLMTest.io, Awesome Agents) have urged caution on vendor-reported numbers. The 'preview' label means some benchmarks haven't been independently replicated. The community notes that while V4 Pro genuinely leads on competitive programming, the SWE-Bench Verified result (80.6%) is effectively a tie with Claude and Gemini — not the 'beats everything' narrative some headlines suggest. Artificial Analysis gave it an Intelligence Index of 52, calling it 'the #2 open weights reasoning model behind only Kimi K2.6.'
**Pricing enthusiasm and concerns:** The pricing ($3.48/M output) generated significant excitement, particularly compared to Claude Opus 4.7's $25/M. However, developers note that V4 Pro uses approximately 190M output tokens to run the Artificial Analysis Intelligence Index — one of the most token-intensive models tested — which partially offsets the low per-token cost. The Flash tier at $0.28/M output has been described as 'the more interesting story' by multiple reviewers.
**Trust and censorship:** The political censorship embedded in training weights remains the most divisive aspect. Enterprise developers in regulated industries note that self-hosting the MIT-licensed weights resolves both the data collection and censorship concerns, but the hardware requirements (~862GB FP4+FP8 for V4 Pro) make this non-trivial. Multiple community discussions reference the geopolitical dimension, with some US-based teams routing only non-sensitive workloads to V4 Pro while keeping Claude or GPT for tasks requiring political neutrality.
**Self-hosting community:** The open-weight release has generated significant interest in the self-hosting community. V4 Flash (~158GB weights) is viewed as the practical self-hosting sweet spot, fitting on 2x H200 or 4x A100 80GB. V4 Pro self-hosting requires enterprise-grade infrastructure (8x H200 node minimum), limiting it to well-resourced organizations. Community quantization efforts are underway but benchmark gaps between quantized and full-precision variants are not yet well characterized.
ユースケース
### 1. High-Volume Code Completion and Review Pipelines
V4 Pro (or more cost-effectively, V4 Flash) is ideal for high-volume code generation pipelines where the model processes thousands of pull requests, generates test suites, or performs automated code review. On LiveCodeBench (93.5%) and SWE-Bench Verified (80.6%), V4 Pro matches or exceeds the best closed models. For a team processing 100K code completions/day, V4 Flash at $0.28/M output tokens costs roughly $3.40/month versus $255/month on Claude Opus 4.7 — a 75x cost reduction with quality within 1-2 points. Choose V4 Pro over alternatives when batch processing volume makes per-token cost the dominant factor and latency is irrelevant.
### 2. Agentic Coding with MCP Tools
V4 Pro's MCPAtlas Public score (73.6%) nearly ties Claude Opus 4.6 (73.8%), making it viable for production agent workflows built on the Model Context Protocol. It ships with pre-tuned adapters for Claude Code, OpenCode, OpenClaw, and CodeBuddy — a developer can swap the base URL and start using V4 Pro in existing agent pipelines within minutes. For a startup building a coding agent product that needs to process multi-step debugging, terminal execution, and tool orchestration at scale, V4 Pro offers near-frontier agentic quality at 1/7th the cost. Choose V4 Pro over GPT-5.4/5.5 for MCP-based agents; choose GPT-5.5 over V4 Pro for long-running autonomous terminal sessions (Terminal-Bench: 82.7% vs 67.9%).
### 3. Long-Document Extraction and Summarization
The 1M-token context window with efficient CSA+HCA attention makes V4 Pro practical for processing entire codebases, legal document sets, or technical research corpora in a single pass. On MRCR 1M, V4 Pro scores 83.5% — ahead of Gemini 3.1 Pro (76.3%) though behind Claude Opus 4.6 (92.9%). For a legal tech company extracting structured data from 500K-token contract sets, V4 Pro handles the context volume at $3.48/M output versus $25/M on Claude. The SMQTS evaluation shows V4 Pro ties or wins 60%+ of pairwise blind comparisons against frontier models on information extraction (N4) and summarization (N2). Choose V4 Pro over alternatives when document volume exceeds what closed-model pricing can sustain and the extraction/summarization quality discount is acceptable.
### 4. Mathematical Reasoning and Scientific Computing
V4 Pro's math strengths are genuine: IMOAnswerBench at 89.8% (ahead of Claude's 75.3% and Gemini's 81.0%), HMMT 2026 at 95.2%, and Apex Shortlist at 90.2%. For a research team using AI to assist with algorithm design, formal verification, numerical analysis, or competitive programming, V4 Pro in Think Max mode provides the strongest open-weight mathematical reasoning available. Choose V4 Pro over Claude for math-heavy workloads; choose GPT-5.4 over V4 Pro for the hardest competition math (HMMT: 97.7% vs 95.2%) if cost is secondary. Note that V4 Pro's mathematical prowess does not always translate to practical software engineering — it can produce mathematically correct code that ignores API constraints or library versioning.
最新ニュース
**April 24, 2026 — V4 Preview Release:** DeepSeek released V4 Pro and V4 Flash as open-weight preview models under the MIT license. Both are available on HuggingFace and ModelScope, with API access via api.deepseek.com. The release includes dual Thinking/Non-Thinking modes, OpenAI and Anthropic API compatibility, and automatic context caching.
**Legacy Model Retirement (July 24, 2026):** DeepSeek announced that legacy model IDs `deepseek-chat` and `deepseek-reasoner` will be fully retired and inaccessible after July 24, 2026 at 15:59 UTC. These currently route to V4 Flash non-thinking/thinking respectively. Migration requires only a one-line model swap; the base URL does not change.
**Promotional Pricing:** V4 Pro launched with a 75% promotional discount ($0.435/M input, $0.87/M output) valid through May 31, 2026. List prices post-promotion are $1.74/M input (cache miss) and $3.48/M output. A 50% off-peak discount applies during Beijing nighttime hours (approx. 11pm–7am Beijing time) on both tiers.
**Third-Party Hosting:** Together AI, Fireworks AI, and Groq are expected to host the open weights with US-based data residency. Together AI pricing starts at $2.00/$4.00 per 1M input/output tokens. Fireworks AI at $1.95/$3.85. Self-hosting estimates range from $0.50-$1.20/M input and $1.20-$2.50/M output on 8x H200 or 4x B200 configurations.
**Regulatory Developments:** Multiple jurisdictions continue to evaluate DeepSeek's data collection practices. Italy, Denmark, Australia, South Korea, and several US states have imposed restrictions or bans. Enterprise users concerned about data residency should consider third-party US-hosted providers or self-hosting the MIT-licensed weights.