One year after its R1 release rattled Silicon Valley, China’s DeepSeek dropped two new flagship models on April 24 that close the frontier gap and obliterate the price floor at the same time.
The two preview releases are DeepSeek-V4-Pro — a 1.6 trillion parameter Mixture-of-Experts model with 49B active parameters — and DeepSeek-V4-Flash at 284B total / 13B active. Both ship with a 1 million-token context window and a new technique the team calls Hybrid Attention Architecture for long-document recall.
Benchmark Numbers That Change the Conversation
- LiveCodeBench Pass@1: V4-Pro-Max scores 93.5 — the highest of any model — beating Gemini 3.1 Pro (91.7) and Claude Opus 4.6 Max (88.8).
- SWE-bench Verified: 80.6% — within 0.2 points of Claude Opus 4.6.
- Codeforces rating: 3,206 — ahead of GPT-5.4 xHigh (3,168).
The Pricing Shock
- V4-Pro: $1.74 input / $3.48 output per million tokens — roughly 1/6 the cost of Claude Opus or GPT-5.5 for comparable benchmark scores.
- V4-Flash: $0.14 / $0.28 per million tokens — undercutting every U.S. frontier lab by an order of magnitude.
Both sets of weights hit Hugging Face the same day under the standard MIT license, putting fresh pressure on OpenAI just 24 hours after its GPT-5.5 unveiling.
Why It Matters
DeepSeek’s V4 release crystalises a now-undeniable shift: open Chinese frontier models are no longer behind on capability and they are an order of magnitude cheaper at inference. For builders, the calculus has changed — for U.S. closed-model labs, the pricing pressure has just become structural.
Follow Vibes Uncut Media for AI model coverage.
Leave a Reply