A year after DeepSeek rattled the AI industry with V3, the Chinese AI lab has released preview versions of its V4 model family — and the numbers are hard to ignore.

DeepSeek V4 comes in two versions: V4-Pro (1.6 trillion total parameters, 49 billion activated per token) and V4-Flash (284 billion total parameters, 13 billion activated). Both ship with a native one-million-token context window. And both are available right now via the DeepSeek API and Hugging Face.

The pricing is where it gets interesting.

V4-Flash costs $0.14 per million input tokens and $0.28 per million output. V4-Pro comes in at $1.74 per million input and $3.48 per million output. For comparison, OpenAI’s GPT-5.5 — announced yesterday — is priced significantly higher. If your workloads run heavy on document analysis, long-context reasoning, or agentic pipelines, those cost differences compound fast.

What’s Actually New Here

The V4 series isn’t just bigger. DeepSeek introduced what it calls Hybrid Attention Architecture, a technique designed to improve how the model handles memory across long conversations. Most LLMs start losing coherence as context grows — the Hybrid Attention approach is DeepSeek’s attempt to fix that at the architectural level.

V4-Pro also ships with a maximum reasoning effort mode, which the company says unlocks stronger performance on knowledge-intensive tasks. Think multi-document synthesis, regulatory research, or complex data queries — the kinds of tasks where a model that can actually hold a million tokens in context makes a real difference.

Both models support the OpenAI ChatCompletions API format and the Anthropic API format, which means if you have existing pipelines built on either platform, migrating to test DeepSeek V4 is a matter of changing an endpoint, not rewriting your code.

The models are open-source. You can run them locally or on your own cloud infrastructure if data sovereignty or compliance is a concern.

How V4-Pro Benchmarks Against the Competition

According to DeepSeek’s own evaluations — and early independent testing — V4-Pro lands close to Gemini 3.1 Pro on world knowledge benchmarks, significantly outperforming other open-source alternatives. On coding and agentic tasks specifically, both V4 variants show strong results, which matters if you’re building workflows where the AI is operating tools rather than just generating text.

Treat benchmark numbers as directional. The more meaningful question is whether the model holds up on your specific tasks at a price point that makes commercial sense.

What This Means for Business

One million tokens is a legitimate enterprise capability. That’s enough to send an entire codebase, a year’s worth of contracts, or a detailed operational dataset as a single prompt. For businesses doing document-heavy work — legal, finance, compliance, procurement — this removes one of the most frustrating limitations of working with AI.

Cost curves are shifting. Twelve months ago, frontier AI capability cost hundreds of dollars per million tokens. V4-Flash at $0.14 per million input tokens is 93% cheaper than the early GPT-4 pricing. That’s not incremental — it changes the math on what’s commercially viable to automate.

Open-source isn’t just for developers. The ability to run V4 on your own infrastructure gives businesses options that hosted APIs don’t: full data privacy, no vendor dependency, and the ability to fine-tune on proprietary data. For regulated industries or enterprises handling sensitive data, that’s a meaningful advantage.

The competition is working in your favour. Every time OpenAI, Google, or Anthropic sees a capable open-source model drop at V4’s price point, it puts downward pressure on their own pricing. The cost of enterprise AI is declining faster than most organisations’ procurement cycles.

The Open Question

DeepSeek is a Chinese AI lab operating under a different regulatory environment than OpenAI or Anthropic. For businesses in sensitive sectors — defence, government, healthcare — that’s a real due diligence consideration, regardless of how good the model is. The open-source release means you can inspect the weights, but the question of where your data goes when you use the API deserves a proper answer before you ship it to production.

For most business use cases, the practical path is to run V4 on your own infrastructure using the open-source weights, which sidesteps the data sovereignty concern entirely.

EDNA’s Take

The story isn’t really about DeepSeek specifically. It’s about what happens when frontier-quality AI becomes a commodity. A model with a 1M-token context window and strong agentic capabilities, available for less than $2 per million tokens, makes the AI cost barrier for most business applications effectively zero.

That shifts the question from “can we afford this?” to “do we have the data, the workflows, and the people to use it well?” Data literacy and operational AI readiness are now the limiting factors — not the price of the model.

If your team is still figuring out how to get structured value from AI, the model race is moving faster than you. The good news is that getting started costs almost nothing. The harder work is building the foundations that make it pay off.

---

Enterprise DNA put together a free field guide on exactly this: the full Claude ecosystem, Claude Code, and how to roll agents out without breaking things. Get the guide.