DeepSeek V4 is out, and it’s a bigger deal than you think

DeepSeek dropped V4 on April 24, and it’s their biggest release since R1 shook the industry back in January 2025. If you remember, R1 was the model that came out of nowhere, trained on limited compute, and suddenly made DeepSeek a household name in AI. Since then, they’ve been quiet—until now.

V4 is a preview, but it already looks like a serious piece of work. The headline feature is a 1 million token context window, which is massive. That means you can throw entire codebases, long documents, or even a full novel at it without breaking a sweat. And unlike some models that choke on long inputs, DeepSeek claims V4 handles it efficiently thanks to a redesigned architecture.

But context length alone isn’t why I’m paying attention. It’s the combination of performance, price, and openness that makes V4 stand out.

Performance that punches above its weight

DeepSeek claims V4-Pro matches or beats closed-source giants like Claude-Opus-4.6, GPT-5.4, and Gemini-3.1 on major benchmarks. That’s a bold claim, but the numbers they shared are convincing. On coding, math, and STEM problems, V4 outperforms other open-source models like Alibaba’s Qwen-3.5 and Z.ai’s GLM-5.1. It also scores well on agentic coding tasks and multistep problem-solving.

They even ran an internal survey of 85 experienced developers—over 90% ranked V4-Pro among their top choices for coding tasks. I’d take that with a grain of salt since it’s their own survey, but it still signals confidence.

Pricing that actually makes sense

V4 comes in two flavors: Pro and Flash. Pro costs $1.74 per million input tokens and $3.48 per million output. Flash is stupid cheap at $0.14 per million input and $0.28 per million output. For context, that’s a fraction of what OpenAI or Anthropic charge for comparable models.

This matters because cost is the real bottleneck for most developers and startups. You can build on V4 without worrying about your API bill exploding. Flash in particular is one of the cheapest top-tier models available, and that’s going to make it a go-to for prototyping and production apps.

Open source done right

DeepSeek kept V4 open source, which is increasingly rare among frontier labs. You can download the weights, modify them, run them on your own hardware. No gating, no API lock-in. This is exactly what the open-source community needs to stay competitive with closed models.

V4 also plays nice with popular agent frameworks like Claude Code, OpenClaw, and CodeBuddy. That’s a smart move—it lowers the friction for developers who already have workflows set up.

The elephant in the room

Let’s be real: DeepSeek hasn’t had an easy year. There have been personnel departures, delayed releases, and growing scrutiny from both US and Chinese governments. The company is a political football now, and that creates uncertainty. Will they face export restrictions? Will future models be censored? These are real concerns.

But for now, V4 is a solid release. It won’t shake the industry the way R1 did, but it doesn’t need to. It’s a practical, well-priced, open-source model that delivers on the fundamentals. And in a world where AI costs keep climbing, that’s refreshing.

If you’re building agentic workflows or just need a capable model that won’t bankrupt you, V4 is worth a serious look. The preview is live now on DeepSeek’s site and API.