DeepSeek V4 is here, and it’s actually interesting

DeepSeek just released a preview of V4, its first flagship model since R1 took the AI world by storm back in January 2025. If you remember, R1 was the one that trained on a shoestring budget and still matched the big guys, turning DeepSeek from a nobody into China’s poster child for AI ambition overnight. That was a wild moment.

V4 isn’t going to repeat that shock value. The industry has moved on, and everyone’s used to Chinese labs dropping competitive open-weight models now. But V4 is still worth paying attention to, for a few solid reasons.

First, it’s genuinely pushing what open-source models can do. DeepSeek claims V4-Pro matches Anthropic’s Claude-Opus-4.6, OpenAI’s GPT-5.4, and Google’s Gemini-3.1 on major benchmarks. That’s not just hype—they published numbers on coding, math, and STEM tasks where V4 beats other open models like Alibaba’s Qwen-3.5 and Z.ai’s GLM-5.1. For developers, that means you can get frontier-level performance without locking yourself into a proprietary API. And the pricing is almost absurd: V4-Pro costs $1.74 per million input tokens and $3.48 per million output tokens. That’s a fraction of what OpenAI or Anthropic charge. V4-Flash is even cheaper at $0.14 per million input tokens and $0.28 per million output tokens. If you’re building applications, this is the kind of pricing that makes you sit up.

Second, the memory architecture is genuinely clever. Both V4 versions handle a 1 million token context window. That’s enough to process entire codebases, long documents, or multi-turn conversations without losing track. DeepSeek says they achieved this through a new design that’s more memory-efficient than the standard attention mechanisms. I’ve seen too many models claim long context and then fail on real-world tasks, but DeepSeek’s internal tests show V4 holds up well on the “needle in a haystack” test—retrieving specific facts from massive texts. If this works in practice, it’s a big deal for agentic workflows and complex reasoning.

Third, DeepSeek is betting big on the agent ecosystem. They optimized V4 specifically for frameworks like Claude Code, OpenClaw, and CodeBuddy. In an internal survey of 85 experienced developers, over 90% ranked V4-Pro among their top choices for coding tasks. That’s not a random number—it suggests real developer buy-in. The model comes in two flavors: V4-Pro for heavy lifting like coding and complex agent tasks, and V4-Flash for fast, cheap inference. Both have reasoning modes that show step-by-step thinking, which is becoming table stakes for serious use.

Now, the elephant in the room: DeepSeek has been through a rough patch since R1. Major personnel departures, delays on previous model launches, and growing scrutiny from both the US and Chinese governments. The company has kept a low profile until this month, when they teased V4 by adding “expert” and “flash” modes to the online version. So V4 is a comeback of sorts, but the geopolitical pressure isn’t going away. It’s unclear how long they can keep releasing open models without running into export controls or domestic restrictions.

Still, V4 is a solid release. It’s not going to upend the industry like R1 did, but it gives developers a real option that’s open, cheap, and competitive. If you’re building on top of AI, that’s exactly what you need.