DeepSeek

Artificial Intelligence · Large Language Models · China Tech

Launch

DeepSeek launches V4-Flash beta, undercutting OpenAI's Luna by 60%

The gains come from retraining, not a bigger model, showing efficiency now beats scale in this pricing fight.

• DeepSeek pushed the post-trained V4-Flash-0731 build into public beta on July 31, 2026, an official release of the model it had previewed back in April.
• DeepSeek is the Chinese AI lab founded by quant-fund veteran Liang Wenfeng out of his hedge fund High-Flyer, known for treating compute and inference cost as design constraints rather than problems money solves.
• The model keeps the same mixture-of-experts architecture as April's preview — 284 billion total parameters, 13 billion active per token, and a 1-million-token context window — with all gains coming from further post-training, not new weights.
• Its Artificial Analysis Intelligence Index score jumped from 40 to 50, nearly matching OpenAI's GPT-5.6 Luna while costing roughly 60% less per task even after Luna's own 80% price cut.
• The launch lands days after OpenAI slashed GPT-5.6 Luna's input price by up to 80% (from $1 to $0.20 per million tokens), turning DeepSeek's move into a direct counterpunch in a three-way price war that also includes Anthropic.
• DeepSeek separately closed a roughly $7.4 billion private funding round in 2026 that pushed its valuation to about $52 billion, with backing from Tencent and NetEase, underscoring the capital base behind its aggressive pricing.
• This shows the AI price war has moved past headline model launches into a grind over inference economics, where post-training efficiency and cache-hit discounts now matter as much as raw benchmark scores.