All The Headlines, All The Model Drops...

DeepSeek launched preview versions of its V4 Pro and V4 Flash models, positioning them as high-performing, low-cost alternatives to US frontier models. These releases undercut competitors like Anthropic's Claude Sonnet 4 by about five times on token pricing—V4 Pro at $1.74 per million input tokens and $3.48 for output, with V4 Flash even cheaper at 14 cents and 28 cents. The models feature a hybrid attention architecture for better long-conversation memory, a one-million-token context window for handling full codebases or documents, and mixture-of-experts scaling to just 37 billion active parameters per task. DeepSeek claims top coding benchmarks and gains in reasoning and agentic tasks, trailing state-of-the-art by three to six months but deployable on cheaper Huawei Ascend chips. Capacity is limited now due to compute shortages, with pricing expected to drop further later this year; the firm is in fundraising talks with Tencent and Alibaba. This builds on DeepSeek's prior R1 model, which shook markets by matching pricier rivals at lower cost, boosting Chinese chip stocks.