DeepSeek released a preview of V4, its new flagship model, on 24 April. It comes in two variants, V4-Pro and V4-Flash, both open-source and available via API.

The release is DeepSeek's most significant since R1, the reasoning model that upended assumptions about Chinese AI capability in January 2025. R1 demonstrated strong performance on limited compute. V4 aims to do the same at a larger scale.

According to the company, V4-Pro matches Anthropic's Claude-Opus-4.6, OpenAI's GPT-5.4, and Google's Gemini-3.1 on major benchmarks. Against open-source peers such as Alibaba's Qwen-3.5 and Z.ai's GLM-5.1, DeepSeek says V4 leads on coding, maths, and STEM tasks.

Pricing is aggressive. V4-Pro costs $1.74 per million input tokens and $3.48 per million output tokens, well below comparable closed-source models. V4-Flash is cheaper still, at roughly $0.14 input and $0.28 output per million tokens.

Both variants support a one-million-token context window, enough to fit the full text of The Lord of the Rings and The Hobbit combined. DeepSeek says this is now the default across all its services.

The detail

The long-context capability rests on an architectural change to the attention mechanism. Rather than treating all prior text equally, V4 compresses older information and concentrates on the most relevant portions. The result, according to DeepSeek's technical report, is sharp efficiency gains.

In a one-million-token context, V4-Pro uses 27% of the compute and 10% of the memory required by V3.2. V4-Flash is leaner still, at 10% of the compute and 7% of the memory. If those figures hold in production, the cost of long-context inference drops substantially.

DeepSeek says it has optimised V4 for popular agent frameworks, including Claude Code, OpenClaw, and CodeBuddy. An internal survey of 85 experienced developers found more than 90% ranked V4-Pro among their top choices for coding tasks.

What this signals

V4 arrives after a turbulent stretch for DeepSeek, marked by personnel departures, delayed launches, and scrutiny from both Washington and Beijing. The release reasserts the firm's position at the front of open-source model development.

For the broader market, the pattern is familiar but still consequential. Each time a capable open-source model closes the gap with proprietary leaders, it compresses the pricing power of closed-source labs and widens the options available to developers building on top of these models.

All benchmark claims cited here are the company's own. Independent evaluations will follow.