DeepSeek’s newest AI model is shaking up the industry. The Chinese AI company has released DeepSeek V4, a family of models that’s competing directly with expensive premium AI systems. And it’s doing it at a fraction of the price.
The V4 family comes in several versions. The biggest is V4-Pro-Max, which uses 1.6 trillion parameters. There’s also a mid-range Pro version and a smaller Flash version with 284 billion parameters. All versions support a massive 1 million token context window, meaning they can process huge amounts of text at once.
The pricing is turning heads. V4 Flash costs just $0.140 per million input tokens. That’s roughly one-third the cost of comparable closed AI models from other companies. Reviewers have rated it 8.3 out of 10 for value.
The benchmarks are hard to ignore. V4-Pro-Max scored 80.6% on SWE Verified, a coding test. It also hit 67.9% on Terminal Bench 2.0 and 83.5% on a long-context memory test called MRCR 1M. On the Apex Shortlist, it scored 90.2%.
Speed is another selling point. V4 Flash averages just 165 seconds per task. That’s faster than several Claude Opus models from Anthropic. V4 Pro averages 256 seconds, which is comparable to Claude Opus 4.7. The Flash version is especially quick on financial tasks, finishing in 18 to 73 seconds.
When compared directly to premium models, DeepSeek V4 Pro tied Claude Opus 4.7 seven-to-seven across 16 financial tasks. It also outperformed Kimi K2.6 in early tests. It does trail GPT-5.4 and Claude Opus 4.7 on the hardest benchmarks by about three to six months.
There are some weaknesses. V4 Pro can be slow on complex multi-turn tasks, sometimes taking over 500 seconds. It’s also very verbose, generating 190 million tokens on one index compared to a median of 42 million for other models.
Still, DeepSeek V4’s combination of strong performance and low cost is making waves across the AI industry. Beyond software applications, the broader energy sector is already exploring how AI can support net zero emissions goals while integrating renewable resources with traditional fossil fuels. Both V4-Pro and V4-Flash are released under the MIT License, allowing developers to freely download, modify, and deploy the open-weight models for local research and commercial applications. DeepSeek V4 Pro received the only perfect 10/10 score in financial research benchmarking for its standout response to a game theory task involving NVDA.
References
- https://fundaai.substack.com/p/deepdeepseek-v4-vs-claude-vs-gpt
- https://www.datacamp.com/blog/deepseek-v4
- https://designforonline.com/ai-models/deepseek-deepseek-v4-flash/
- https://artificialanalysis.ai/models/deepseek-v4-pro
- https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro
- https://mashable.com/article/deepseek-v4-preview-comparison-chatgpt-claude-gemini
- https://llm-stats.com/models/deepseek-v4-pro-max
- https://www.youtube.com/watch?v=eV3lAY77IpU
