DeepSeek released V4-Flash on Friday, and research firm Artificial Analysis says it is the least expensive well-known model to run globally, scoring 50/100 on their index at $0.28 per million output tokens versus Claude Opus 4.8 at $25. The model is not the highest-performing but wins decisively on cost, intensifying pressure on US AI firms.
Alibaba unveiled its most capable model yet, built on the Qwen 3.5 architecture with 2.4 trillion parameters and 95 billion active parameters. The company says it delivers comparable or better scores than Anthropic's Fable 5 on several benchmarks, and confirmed open weights will be released next week. Alibaba shares jumped 6% on the news.
Article 50 obligations now require chatbot disclosure, machine-readable marking of synthetic content, and labeling of deepfakes and AI-written text. The rules apply to every company calling a foundation model API in the EU, and vendors cannot comply on behalf of deployers.
OpenAI reduced Luna prices by 80% and Terra by 20%, citing GPU kernel improvements and better speculative decoding as Chinese models drive a price war. Separately, OpenAI revealed an unreleased model called Astra that solved 10 major math problems, while Anthropic claims its Fable model solved 5 of them.
A Defense One report highlights how the future of AI is increasingly driven by startups using open-source software to build cheaper alternatives, but security threats are rising in parallel. Hugging Face's CEO called for mandatory disclosure laws after the OpenAI breach, arguing that restricting powerful models is not the solution.
Amazon, Google, Meta, and Microsoft have collectively spent over $1 trillion on AI infrastructure since 2023, including data centers, chips, and power. The spending is straining power grids, capital, and construction capacity, with the inference market projected to double the size of training in coming years.