On April 24, 2026, DeepSeek launched V4 and once again shook up the AI market. With a model of 1.6 trillion parameters, a Mixture of Experts architecture, and an entry price of just $0.14 per million tokens, the Chinese company repeats the play that in 2025 surprised the world with DeepSeek R1: delivering frontier capabilities at a fraction of the cost of paid models.

Here we explain what DeepSeek V4 is, how it compares with GPT and Claude, and what this launch means for AI developers and professionals in Latin America.

What DeepSeek V4 is and what its variants are

DeepSeek V4 comes in two versions with different architectures:

• DeepSeek V4 Pro: the full model with 1.6 trillion total parameters in a Mixture of Experts (MoE) architecture. It only activates a fraction of the parameters per query, which makes it more computationally efficient than the total number suggests. Context of up to 1 million tokens.

• DeepSeek V4 Flash: a more compact version of 284 billion parameters, designed for speed and cost. Price of $0.14 per million input tokens, against the $2.50 of GPT-4o Mini and the $3.00 of Claude Haiku 3.5.

The 1-million-token context window in the Pro version is especially relevant for use cases that require analyzing extensive documents, complete codebases, or long conversations without losing context.

How it compares with GPT and Claude in benchmarks

According to the benchmarks published by DeepSeek and analyzed by Bloomberg and TechCrunch, V4 Pro surpasses GPT-4o in mathematical reasoning (MATH-500) and positions itself competitively against Claude Opus in code comprehension (HumanEval).

In general reasoning tasks and following complex instructions, the model shows results comparable to the paid models of OpenAI and Anthropic, although with differences in low-resource languages and in the quality of responses in Latin American Spanish, where Western models still have an advantage.

The real differentiator is not technical: it's economic. For high-volume applications where the cost per token matters, DeepSeek V4 Flash completely changes the viability analysis.

MoE architecture: why it matters

The Mixture of Experts architecture divides the model into specialized "experts", and for each query it only activates a subset of them. This makes it possible to have a very large model in terms of total parameters, but with a much lower inference computational cost than an equivalent dense model.

DeepSeek already used this architecture in previous versions, but V4 scales it significantly and combines it with a 1-million-token context window, which opens up use cases that previously required very costly infrastructure or complex chunking pipelines.

Open source: what it means and how to access it

DeepSeek V4 is published under an open source license that allows you to use the model, modify it, and deploy it on your own infrastructure. This fundamentally differentiates it from GPT and Claude, which are only accessible via API with rates controlled by OpenAI and Anthropic respectively.

You can access it in three ways:

• DeepSeek API: the simplest option. You create an account at platform.deepseek.com, get an API key, and start making calls. The interface is compatible with the OpenAI API format, which makes it easy to migrate existing projects.

• Hugging Face: the model is available on the Hugging Face hub for download and local use or on your own infrastructure.

• Ollama: to run it locally on your computer (requires a GPU with enough VRAM for the smaller versions).

If you're interested in comparing AI models and understanding when to use each one, also read our article about GPT-5.4-Cyber and OpenAI's specialized models, which rounds out the picture of the current ecosystem.

Implications for the open source ecosystem

Every DeepSeek launch generates the same debate: can open source models match —or surpass— the closed models of the big Western companies? V4's answer suggests they can, at least in standard benchmarks and at a fraction of the cost.

For the developer ecosystem in LATAM, this is good news. Access to frontier models no longer depends exclusively on having a budget for costly APIs. DeepSeek V4 Flash, in particular, makes it viable to build sophisticated AI applications at very low costs, which democratizes access to the technology for startups, freelancers, and small teams.

Recommended Coderhouse courses

If you want to learn to integrate models like DeepSeek V4 into your projects and understand the frontier AI ecosystem, these courses give you the foundations and the practical tools:

• Introduction to Artificial Intelligence: understand how LLMs work, the differences between models, and when to use each one.

• AI Engineering: learn to build applications with multiple LLM providers, including open source models like DeepSeek.

• AI Agents: develop autonomous agents that take advantage of the extended context capabilities of models like DeepSeek V4 Pro.

Frequently asked questions

Is DeepSeek V4 really comparable to GPT-4o?

In public benchmarks for math and code, V4 Pro shows results competitive with GPT-4o. However, benchmarks don't capture everything: the quality in languages other than English, following complex instructions, and multi-step reasoning capacity in real-world contexts may vary. The most advisable thing is to test it with your specific use cases before migrating a complete project.

Is it safe to use DeepSeek V4 in enterprise projects?

It depends on the use case. If you process sensitive data, the open source option allows you to deploy the model on your own infrastructure without sending information to external servers, which can meet stricter privacy requirements. For public or non-sensitive data, the API is a viable and economical option.

What does it mean that it's open source?

It means the model's code and weights are public: you can download them, modify them, and use them without depending on DeepSeek's servers. Unlike GPT and Claude, which only exist as APIs, DeepSeek V4 can be run on your own infrastructure.

How does this launch affect OpenAI and Anthropic?

Every competitive low-cost model pressures OpenAI and Anthropic to lower prices or improve the cost-performance ratio. It already happened with DeepSeek R1 in 2025, which accelerated price cuts in the OpenAI API. It's expected that V4 will generate similar pressure. For developers, that's always good news.

Does it work well in Spanish?

DeepSeek V4 works in Spanish, but the quality is variable compared to English. For Latin American Spanish applications that require high linguistic precision (copywriting, customer service, editorial content), Anthropic's or OpenAI's models still have an advantage. For technical cases like code generation or data analysis where language is less critical, the difference is reduced considerably.