DeepSeek V4 released, can its programming capabilities beat GPT and Claude? Costs dominate the leaderboard again

📄Full Article· Automatically extracted by trafilaturaGemini 翻譯1167 words

DeepSeek V4 Preview was officially released and open-sourced on the 24th, launching two models: V4-Pro (1.6 trillion parameters, MoE architecture) and V4-Flash (284 billion parameters), both supporting a 1 million token context window. (Previous coverage: DeepSeek V4 rejects NVIDIA, turns to Huawei! Alibaba, ByteDance, and Tencent rush to buy Ascend 950 PR chips) (Background: DeepSeek V4 released in late April, Liang Wenfeng internally confirms: 1 million token window, runs entirely on Huawei Ascend 950 PR chips) Amidst long-standing community anticipation, DeepSeek officially launched its latest open-source large model, DeepSeek V4 Preview, today (24th), making a 1 million token context window a standard feature across the entire series. Is it enough to match the current closed-source models from Silicon Valley? In the SWE-bench code test, DeepSeek V4 Preview surpassed Claude Opus 4.6 (80.9%) and GPT-5.2 (80%) with a score of 83.7%. While it has not yet caught up to the latest flagship models, it is already impressive enough. The V4 series launched two models: V4-Pro (1.6 trillion parameters, 49 billion active per inference) and V4-Flash (284 billion parameters, 13 billion active). The architecture is Mixture-of-Experts (MoE). Simply put, the model does not mobilize all neurons every time, but calls on a small subset of "experts" as needed, significantly compressing computational consumption without sacrificing response quality. The context window reaches 1 million tokens. It can read the entire seven-book "Harry Potter" series plus a massive codebase at once. DeepSeek-V4-Pro Max is not "sweeping all leaderboards," but its positioning is clear: it is very strong in reasoning, coding, and Agentic tasks. - In knowledge reasoning, it is close to GPT-5.4 / Gemini-3.1-Pro / Opus-4.6 - Coding and Agentic capabilities are particularly outstanding - However, it is still outperformed by GPT-5.4, Gemini-3.1-Pro, or Opus-4.6 on some high-difficulty reasoning leaderboards. The Agentic performance of DeepSeek-V4-Pro Max can be described as steadily strong: - It scored 80.6 on SWE Verified, almost equal to Gemini-3.1-Pro and slightly lower than Opus-4.6 - It scored 83.4 on BrowseComp, also close to Opus-4.6's 83.7, trailing only Gemini's 85.9 - It scored 73.6 on MCPAtlas Public, almost tying Opus-4.6's 73.8, leading both GPT-5.4 and Gemini DeepSeek V4 is priced at approximately $0.30 per million input tokens, which is 27 times cheaper than comparable closed-source models. Meanwhile, V4 runs entirely on Huawei Ascend 950 PR chips without using NVIDIA CUDA. Against the backdrop of tightening U.S. export controls, this is not just an engineering choice but a long-term survival strategy, and DeepSeek has provided the first large-scale viable version with V4.

Data Status✓ Full text extractedRead Original (動區 BlockTempo)

🔍Historical Similar Events· Keyword + Asset Matching6 items

2026-05-03

GPT-5.5 security testing scores are nearly identical to Claude Mythos, with the latest evaluation debunking Anthropic's claim that it is too dangerous to be released.

Similarity 140%關鍵字 claude/gpt同分類 zh

2026-05-03

GPT-5.5 security testing scores are nearly identical to Claude Mythos, debunking Anthropic's claim that it is too dangerous to be released.

Similarity 140%關鍵字 claude/gpt同分類 zh

2026-04-24

DeepSeek V4 Is Here—Its Pro Version Costs 98% Less Than GPT 5.5 Pro

Similarity 130%關鍵字 gpt/deepseek

2026-04-23