DeepSeek V4 Preview Drops: 1.6T Parameters, 1M Context, Apache 2.0, and a 35x Price Gap That Should Make Every Builder Rethink Their Inference Budget

DeepSeek released V4 Preview on April 24 with two open-weight variants: V4-Pro (1.6T total parameters, 49B activated via MoE) and V4-Flash (284B parameters, 13B activated). Both support 1M-token context windows. Both are Apache 2.0 licensed. Both are live right now on Hugging Face, the DeepSeek API, and chat.deepseek.com.

The efficiency numbers are what caught my attention. V4 achieves 27% of V3.2's FLOPs and uses only 10% of the KV cache through a new Hybrid Attention Architecture. That's not an incremental improvement. That's a fundamentally different cost curve for inference. And the pricing reflects it: $0.14 per million input tokens on the DeepSeek API, versus GPT-5.5's $5 per million. A 35x spread for models that compete within a few points on the same benchmarks.

Simon Willison's analysis nailed the framing: "almost on the frontier, a fraction of the price." V4-Pro scores 80.6% on SWE-Bench Verified. GPT-5.5 scores 88.7%. That 8-point gap is real, but for the vast majority of production workloads, 80.6% is more than enough. Especially when your inference bill drops by an order of magnitude.

The timing isn't a coincidence. DeepSeek shipped this the same day as GPT-5.5's launch, turning what should have been OpenAI's victory lap into a pricing comparison. Meanwhile, Tencent and Alibaba are reportedly in talks to invest in DeepSeek at a $20B+ valuation, which has roughly doubled from initial targets. DeepSeek also confirmed Huawei-based inference is coming: 950 supernodes launching in H2 2026 with significant price drops expected once the full cluster is online. That's the first major frontier-competitive model running production inference entirely on non-NVIDIA hardware.

The technical report confirms no multimodality in the current release. Community consensus is this is deliberate, not a limitation. While competitors offer native vision, DeepSeek bet on depth over breadth at 1.6T parameters.

For builders, the action item is straightforward: if you're running inference workloads where 80% SWE-Bench accuracy is good enough (and for most production code, it is), you should be evaluating V4-Pro today. Run your actual prompts through it. Compare output quality on your specific use case, not on benchmarks. The 35x cost difference means the ROI calculation isn't close for many workloads. I'm planning to test it against my own RAG pipeline this weekend.