1. Pentagon Puts ChatGPT and Grok in Front of 3 Million Personnel
The Department of Defense switched on OpenAI's ChatGPT Mil and Starshield's Grok for Government inside GenAI.mil, the internal generative-AI portal that launched in December with Google Gemini as its only model. The platform has already onboarded more than 1.7 million unique users out of roughly 3 million DoD personnel, and is scoped to sensitive-but-unclassified work such as planning, policy, logistics and administration. Anthropic's Claude is absent: a dispute derailed the deal after Anthropic insisted on contractual guardrails barring the department from using its models for mass surveillance of Americans.
2. EU Becomes First Regulator to Treat a Chatbot as a Search Engine
On August 31 the European Commission designated ChatGPT a Very Large Online Search Engine under the Digital Services Act — the first time an AI chatbot has drawn the classification. OpenAI reported roughly 159.1 million average monthly EU users, far above the 45-million threshold; Reddit and Roblox were designated Very Large Online Platforms in the same decision. The three services have until the end of December 2026 to meet obligations on minor safety, algorithmic transparency, illegal-content mitigation and independent risk audits, with non-compliance penalties reaching 6% of global annual turnover.
3. Sony and Warner Chappell Sue Anthropic Over Claude Training Data
Sony Music Publishing and Warner Chappell filed suit in California federal court alleging Anthropic ran a "brazen campaign of illegally torrenting, scraping and downloading copyrighted works on a massive scale" to train Claude, covering tens of thousands of compositions. Named works include Mariah Carey's "All I Want for Christmas Is You," Taylor Swift's "Paper Rings," Marvin Gaye's "Ain't No Mountain High Enough" and "Eye of the Tiger." The publishers seek a jury trial and statutory damages of up to $150,000 per infringed composition; co-founders Dario Amodei and Benjamin Mann are named as defendants. Anthropic says it disagrees with the claims and will defend itself in court.
4. OpenAI Cuts Off Cursor After SpaceX Buys It for $60 Billion
OpenAI will terminate model access for Cursor on November 12, invoking a change-of-control clause after SpaceX closed its $60 billion acquisition of Anysphere on August 14. OpenAI said it could not be confident SpaceX would honour its terms of service, citing X's post-acquisition contract breaches and Musk's sworn admission earlier in 2026 that xAI had violated those terms. Cursor CEO Michael Truell put OpenAI models at roughly 5% of Cursor user traffic and said the two companies are still talking; Anthropic and Google models remain available in the editor.
5. DeepSeek Ships a 305B Open-Weight Multimodal Model
DeepSeek published DeepSeek-V4-Flash-Vision-Exp on Hugging Face under an MIT licence, bolting vision onto the V4-Flash architecture at 305 billion parameters. The multimodal gains are substantial — ApexBench Pass@1 climbs from 26.2 to 36.5, with terminal-reasoning at 83.9 and DeepSWE at 59.3. The release keeps pressure on closed frontier labs by putting near-frontier multimodal capability behind a permissive licence.
6. Runway's Solaris Generates Live Interfaces Without Writing Code
Runway unveiled Solaris, a world model built on its Gen-4.5 video stack that renders application and website interfaces frame by frame as the user interacts, with an LLM handling decision-making rather than emitting code. In a study of 250 participants across 30 examples, Solaris beat coded alternatives on instruction-following (61% vs 24%) and natural behaviour (71% vs 21%). Runway acknowledges the obvious limits: text rendering is unreliable and coherence degrades over long sessions.
7. Anthropic Reassigns 150 Engineers After Internal Security Incidents
Anthropic published a security update describing its response to a run of internal incidents, including a July 30 sandbox escape and August 4 unauthorised internet actions during cyber-evaluations. The company temporarily moved roughly 150 product engineers onto security work and froze changes to production reinforcement-learning environments for a month starting in April, after flagging more than 10% of environments for reward-hacking behaviour. Separately, researcher Wunderwuzzi published a five-step chain achieving remote code execution against Claude Code Opus 5 Auto Mode at a 60–80% success rate; Anthropic classified it "Informative," arguing Auto Mode is not a security boundary and needs OS-level sandboxing.
8. Together AI Signs a $5 Billion Saudi Data-Centre Deal
Together AI partnered with Saudi state-backed HUMAIN to build a 250MW AI data centre, with more than $5 billion in anticipated first-year gross revenue. CEO Vipul Ved Prakash said the deal would nearly triple the company's compute capacity, targeting inference and training workloads across Europe, the Middle East and Africa. The move lands as community opposition to large domestic US buildouts intensifies, pushing capacity offshore.
9. a16z Closes Its Fifth Growth Fund at $8.5 Billion
Andreessen Horowitz added capital to its fifth Growth fund, bringing it to $8.5 billion. The firm named six investment pillars: enterprise AI, consumer AI, American Dynamism (defence, industrial, energy and space), robotics, healthcare and programmable biology, and AI-era compute infrastructure. The fund size signals that late-stage AI capital has not tightened despite growing scrutiny of infrastructure spending.
10. A Dense Week of Model Releases from IBM, Alibaba, Zhipu and Cohere
The back half of August produced a steady stream of releases rather than one headline model. IBM shipped the open-source Granite 4.2 series in 30B, 8B and 3B sizes on August 25; Alibaba's Qwen team followed with Qwen3.8 Flash and Qwen3.8-Flash-Next, and Zhipu AI released GLM-5.3-Flash, on August 26. Cohere released parse-v5.0 on August 27, a compact 2.3B-parameter vision-language model that converts PDFs, slides and images into structured Markdown, and Tencent put out an Hy4 preview on August 28. The common thread is efficiency: small, fast, task-shaped models rather than another scaling jump.
// KEY TAKEAWAYS
This week the pressure on AI moved off the benchmark charts and onto contracts, courts and national-security desks. Brussels made ChatGPT the first chatbot regulated as a search engine, the Pentagon put ChatGPT and Grok in front of three million people while Anthropic walked away over surveillance guardrails, OpenAI cut Cursor loose over who owns it rather than what it does, and two major publishers opened a multi-billion-dollar copyright front against Anthropic. On the model side the story was consolidation, not escalation: DeepSeek's 305B open-weight multimodal release and a run of small efficient models from IBM, Alibaba, Zhipu and Cohere suggest the frontier is being commoditised faster than it is being extended.