Issue No. 4Oct 19 – 25, 2026Week 43
The 10 AI stories that mattered this week
DeepSeek faced sweeping US federal restrictions, while Stripe agreed to buy OpenRouter for $7.5B and Anthropic disclosed a draft plan for more than $600B in compute spending. New models also pushed speed and autonomy, as OpenAI resumed training in an offline environment and Recursive released an AI-built coding model.
01 · Regulation · techmeme.com
The US DHS ordered all federal civilian agencies to remove DeepSeek technology and barred future purchases, expanding government restrictions on the Chinese AI company.
02 · Industry · startupfortune.com
Stripe agreed to acquire AI model router OpenRouter for $7.5B, outbidding Databricks and paying six times its May valuation in a major AI deal.
03 · Models · huggingnews.com
Kimi launched its K3.1 Pro flagship AI model for professional work, with the newest version of its assistant now rolling out across the service.
04 · Models · inceptionlabs.ai
Inception’s Mercury 3 is billed as the most capable diffusion LLM yet, outperforming GPT-6 Luna and Claude Haiku 5.5 while running 4–8x faster at 1,000+ tokens per second.
05 · Funding · marketbrief.now
A draft Anthropic S-1 says revenue reached $16.3B in H1 2026 and outlines over $600B in compute spending, including $450B in noncancelable contracts over 7–10 years.
06 · Regulation · madrobot.blog
The US ordered every civilian agency to remove DeepSeek and third-party tools built on it, and to stop buying them, broadening the ban beyond DeepSeek’s own products.
07 · Tools · techcrunch.com
Recursive released Flint, a free laptop-sized coding model built largely by its own AI research system, offering an early test of AI that improves itself.
08 · Infra · huggingnews.com
OpenAI resumed frontier model training and internal inference on Oct. 1 in a cached offline-world environment designed to prevent unauthorized network breaches after a Sept. 20 incident.
09 · Agents · wired.com
About 80 WhatsApp help pages exposed hidden instructions for Meta’s chatbot, including “WHAT NOT TO SAY”; Meta said a bug caused the leak.
10 · Models · techmeme.com
Inception launched Mercury 3, a diffusion-based LLM it claims is comparable to Luna and Haiku models but 4–8x faster, generating about 1,000 tokens per second.