1. OpenAI Says Astra Is the First Model to Cross Its ‘Critical’ Cyber Threshold
OpenAI announced on September 1 that its Astra model meets the ‘Critical’ cybersecurity capability tier of its Preparedness Framework — the first model it has ever designated at that level. In expert-led evaluations, Astra found previously unknown vulnerabilities in a hardened browser and OS, chained them into a working sandbox escape that executed commands on the host, and built a local privilege-escalation chain from an unprivileged user to root. OpenAI is gating the release, restricting access to the most powerful cyber capabilities while it hardens safeguards.
2. Anthropic Ships Claude Fable 5.1 and Mythos 5.1
Anthropic released Fable 5.1 publicly and Mythos 5.1 under trusted access on September 1. Headline pricing is unchanged at $10/M input and $50/M output, but cache reads dropped roughly 75% to $0.25/M, making cache-heavy workloads about 25% cheaper. The biggest reported jump is on Terminal-Bench-Science, where Fable 5.1 scores 52.6 against Fable 5's 24.7, alongside 73.4% on CursorBench 3.2.0.
3. Sony Music and Warner Chappell Sue Anthropic Over Song Lyrics
The two publishers filed in the Northern District of California on August 28, accusing Anthropic of large-scale scraping and torrenting of ‘tens of thousands’ of copyrighted compositions to train Claude — including Hallelujah, Uptown Funk, Eye of the Tiger and Mariah Carey's All I Want for Christmas Is You. They seek up to $150,000 per infringed work plus $25,000 per removal of copyright management information, and named co-founders Dario Amodei and Benjamin Mann as personal defendants. Anthropic disputes the allegations and says it will defend itself in court.
4. Anthropic Signs a $35B Cloud Deal With Nvidia-Backed Lambda
Signed August 31, the agreement covers roughly 350 megawatts of training and inference capacity at a Hut 8 data center in Nueces County, Texas. Nvidia plays three roles at once — chip supplier, Lambda equity backer, and lease anchor on the 700MW campus — which has drawn fresh scrutiny of circular financing in AI infrastructure. It lands days after a separate $45B commitment from Anthropic to rent capacity from Nscale in West Virginia.
5. The Pentagon Adds ChatGPT Mil and Grok to GenAI.mil
Custom builds of OpenAI's ChatGPT and xAI's Grok joined Google Gemini on the Department of Defense's central generative-AI portal on August 31. The platform gives DoD staff frontier models without routing sensitive government data through consumer channels, and is exempt from the usual consumer data collection. GenAI.mil has onboarded more than 1.7 million unique users out of about 3 million DoD personnel.
6. ChatGPT Health Connects to Epic Health Records
OpenAI launched a read-only Epic EHR integration on September 1, letting authorized clinicians pull appointment notes, lab results, medications and specialist documentation into ChatGPT or work with it inside Epic. The model cannot write back to patient records. OpenAI reports physicians rated 99.1% of responses safe across 4,363 ratings covering 27 clinical use cases; launch partners include UCSF, Cedars-Sinai, Memorial Sloan Kettering, Boston Children's Hospital and HCA Healthcare.
7. World Labs Unveils Atlas, a Spatial Intelligence Model
Atlas is a multimodal transformer that generates up to a minute of 1440p video and reconstructs 3D scenes from as few as one to three images. World Labs reports it wins 81–93% of head-to-head comparisons against specialized video baselines while beating specialist models on the DTU, ETH3D, KITTI and ScanNet benchmarks — an argument that a single general spatial model can outrun purpose-built pipelines.
8. Dell Books a Record $60.9B in AI Orders
Dell's Q2 FY27 revenue hit $46.97B, up 58% year over year, with AI server revenue of $16.4B and record AI orders of $60.9B pushing backlog to $95B. Management raised full-year revenue guidance by $25B to $192B. The numbers are one of the clearest reads yet on how much of the AI buildout is landing on traditional server vendors rather than the hyperscalers.
9. New York City Bans Student-Facing AI Chatbots Through 8th Grade
The NYC Department of Education released a policy prohibiting student-facing generative AI — chatbots and AI tutors included — from pre-K through 8th grade. The same policy caps device screen time at 30 minutes for grades 3–5 and 45 minutes for grades 6–8. As the largest school district in the US, NYC's line is likely to become a reference point for districts still writing their own rules.
// KEY TAKEAWAYS
Two themes dominated the week. First, capability is outrunning the release process: OpenAI declared Astra ‘Critical’ on cyber and gated it, while Anthropic and World Labs pushed big benchmark jumps out the door — the frontier is now shipping with safety brakes attached rather than after them. Second, the bill is coming due on both sides of the ledger: Anthropic committed $80B across two compute deals in a single week while simultaneously being sued by Sony and Warner over the data that built its models. Meanwhile deployment kept widening at the edges — 1.7M Pentagon users, Epic health records, Dell's $95B backlog — even as New York City drew the first hard line the other way.