Over the past two years, a Chinese company has been quietly reshaping the landscape of artificial intelligence. While OpenAI, Google, and Anthropic compete to see who can charge the most per token, DeepSeek has taken the opposite approach: delivering top performance at irrelevant prices.
The result? In May 2026, DeepSeek launched the V4 Preview — two models (Pro and Flash) that not only rival the best closed models in the world but also cost 10x to 50x less and are completely open source.
How Does DeepSeek Manage to Be So Much Cheaper?
The story of DeepSeek's innovation is a lesson in efficiency. Each generation has brought a leap in cost-effectiveness:
DeepSeek-V2 (2024)
The first major impact. Introduced the MoE (Mixture of Experts) architecture with selective parameter activation. While traditional models activate 100% of the parameters for each request, DeepSeek activates only the necessary "experts" for each task. The result: large model performance at small model cost.
DeepSeek-V3 (2025)
The game changer. With 671B total parameters (37B active), V3 was trained for US$ 5.5 million — compared to hundreds of millions of dollars for equivalent models. The market took notice: it was no longer acceptable to spend US$ 200 million to train a model when it could be done for 3% of that.
DeepSeek-R1 (2025)
The reality check. While the world applauded OpenAI's o1 as a "revolution in reasoning", DeepSeek launched R1 with similar chain-of-thought (CoT) capabilities — for a fraction of the cost. The announcement sent tech stocks tumbling in the American market. Silicon Valley realized that the AI monopoly was nearing its end.
The Technical Innovations That Enable Low Costs
It's not magic — it's engineering. DeepSeek developed three key innovations:
1. Ultra-Efficient MoE Architecture
The V4-Pro has 1.6 trillion total parameters, but activates only 49 billion per inference — a mere 3%. It's like having 1,600 specialists on the team, but calling only 49 for each task. The rest remain available but do not consume resources.
2. Optimized Training
Proprietary parallelization and optimization techniques allow for training giant models with much less computational power. The estimate is that V4-Pro cost US$ 12 million to train. Comparison:
| Model | Training Cost | Difference |
|---|---|---|
| DeepSeek V4-Pro | ~US$ 12 million | — |
| GPT-5 | ~US$ 200+ million | ~17x more expensive |
| Gemini 3 | ~US$ 300+ million | ~25x more expensive |
3. Open Source = No Artificial Margin
Unlike American competitors that need to generate returns for investors, DeepSeek operates with much tighter margins. The model is open — anyone can download, inspect, modify, and run it locally. The API exists as a convenience, not as the only option.
The Result: Costs 10x to 50x Lower
In practice, this means that your company can use cutting-edge AI for pennies. See the cost comparison for inference:
| Model | Cost per 1M tokens (input) | Cost per 1M tokens (output) |
|---|---|---|
| DeepSeek V4-Flash | ~US$ 0.05 | ~US$ 0.15 |
| DeepSeek V4-Pro | ~US$ 0.20 | ~US$ 0.50 |
| GPT-5 | US$ 2.50 | US$ 10.00 |
| Claude 4 (Opus) | US$ 3.00 | US$ 15.00 |
| Gemini 3 Ultra | US$ 1.50 | US$ 5.00 |
The V4-Flash costs 50x less than Claude 4 to generate text. This is not a margin optimization — it's a category change.
And it was exactly this trajectory of innovation that led to the announcement we are about to explore.
🚀 DeepSeek-V4 Preview: The Announcement
On May 16, 2026, DeepSeek released to the public the DeepSeek-V4 Preview. It's not just another model — it's the culmination of everything the company has been building: elite performance, ultra-low cost, and open source.
The V4 Preview comes in two versions:
| Specification | V4-Pro | V4-Flash |
|---|---|---|
| Total parameters | 1.6 trillion | 284 billion |
| Active parameters | 49 billion | 13 billion |
| Pre-trained tokens | 33 trillion | 32 trillion |
| Maximum context | 1 million | 1 million |
| Open source | ✅ | ✅ |
| API available | ✅ | ✅ |
| Profile | Expert (precision) | Instant (speed) |
DeepSeek-V4-Pro: Raw Power
With 1.6 trillion parameters and performance that DeepSeek claims rivals the best closed models in the world, the V4-Pro is for tasks that require the highest level of precision and sophistication.
DeepSeek-V4-Flash: Speed and Economy
With 284B total parameters and only 13B active, the Flash is the everyday model. Fast, cheap (US$ 0.05/1M input), and surprisingly capable. Ideal for chatbots, quick analyses, and large-scale automations.
1 Million Tokens of Context: What Changes
The context of 1 million tokens (about 750,000 words) allows for processing at once:
- 📚 Complete books — analyze entire works without splitting into chapters
- 💻 Entire codebases — understand complete software projects
- 📄 Hundreds of documents — extract insights from PDFs simultaneously
- 💬 Endless conversations — chatbots that remember everything
- 🏛️ Contracts and manuals — complete legal documents
No other open model offers this level of context with consistent quality.
How This Empowers Businesses
The combination of low cost + 1M context + open source changes the math for various segments:
Scale Customer Service Automation
Chatbots with 1M token context maintain long conversations without losing history. At US$ 0.05 per 1M tokens, the cost is practically zero. Customer support with senior-level quality for pennies.
Legal and Document Analysis
Input entire contracts and receive complete analyses: abusive clauses, risks, suggestions. In seconds. Medium-sized firms can automate 70% of document review.
Financial Processing
Quarterly reports, balance sheets, income statements — all processed in batches. Extract indicators, generate executive summaries, detect anomalies automatically.
Specialized Technical Support
Load manuals, FAQs, and complete documentation into context. Provide high-level technical support without training a model. Great for software, equipment, and healthcare companies.
Education and Corporate Training
Complete educational materials in context allow AI tutors to master the content from start to finish. Personalized training for each employee.
Open Source = Data Sovereignty
For Brazilian companies, the fact that DeepSeek-V4 is open source has an extra benefit: you can run the model on your own infrastructure in Brazil, without relying on external APIs and without sending data outside the country. LGPD-friendly.
Official Links
Conclusion
DeepSeek's trajectory proves that cutting-edge artificial intelligence doesn't have to be expensive. From V2 to V4, the Chinese company has innovated in architecture, training efficiency, and business model to deliver elite performance at affordable prices.
The DeepSeek-V4 Preview is the pinnacle of this journey. With 1 million tokens of context, costs 10x-50x lower than competitors, and open source, it is not just another release — it is the consolidation of a new era in artificial intelligence.
For Brazilian companies, the message is clear: the technology to transform your business with AI is available, accessible, and within everyone's reach. The question is no longer "if" — it's "how" to do it.
👉 Want to implement generative AI in your business? Talk to Agência Kaizen.