Anthropic told affected users that a bad actor is lifting Claude login sessions off their computers with common infostealer malware, then spending the stolen accounts' paid tokens — in one case minting unauthorized Claude Code OAuth tokens from a compromised session key. Anthropic signed users out, invalidated authorizations, and issued partial refunds, but account support tracks only total usage, never an itemized breakdown. That detection gap is the developer-side problem: a machine-local infostealer defeats every server-side control, and no tool exists to show what is consuming an allowance, so theft can run for months unnoticed.

OpenAI released ChatGPT Images 2.5 five months after Images 2, cutting generation latency by up to 50% and adding Sketch for drawing references, format templates, and prompt sharing. Two API models ship alongside it — GPT-Image-2.5 Flare as the default and Sunburst for tighter control across edits — so developers building visual products now pick between throughput and edit fidelity rather than settling for one image model. The speed gain is the concrete change; the rest is packaging around an incremental quality bump.

Meta's new US personal AI agent connects to email, calendar, payments, health, and smart-home apps to send emails, book travel, and check out via Link by Stripe, running on web, iOS, Android, and WhatsApp — free, with $20 Power and $100 Maximum tiers. Meta says Muse executes inside a dedicated secure VM with its own browser, watched by a system-isolated Sentinel agent, and that its data never reaches Meta's ads systems; those claims land two weeks after an $18B multistate settlement and await scrutiny from security researchers.

ASML and TSMC opened an industry initiative to move High NA EUV lithography from 6-inch to 12-inch photomasks, targeting a mask pilot line in 2031 and production-ready systems by 2033 — a roadmap commitment, not a shipping capability. TSMC still starts High NA high-volume manufacturing on 6-inch masks in 2030, so the near-term leading-edge economics are unchanged; the larger format is where ASML's Christophe Fouquet expects scanner productivity gains and the end of stitching constraints to show up, and mask suppliers across the ecosystem now have a dated target to build toward.

Google Cloud and Accenture are standing up the Accenture Gemini Enterprise Business Group, with Google training up to 1,000 Accenture forward-deployed engineers to build custom applications on Gemini Enterprise inside customer organizations. Ramp's August data puts Google at roughly 6% of US enterprise AI spending among its customers against Anthropic's 43.5% and OpenAI's 39.7% — a gap Google disputes as missing large strategic deals — and consulting headcount is now the channel hyperscalers are buying to convert $811B in Alphabet purchase commitments into enterprise revenue.

The Department of Energy approved a $1.9 billion loan to NextEra Energy to restart the 615-megawatt Duane Arnold plant in Iowa, mothballed since 2020 storm damage, with power targeted for 2029 and Google reportedly planning up to six data centers nearby. It is the second such federal loan after $1 billion to Constellation for Three Mile Island, marking restarted nuclear as the administration's preferred answer to AI power demand. Only 50 megawatts is reserved for the local cooperative — 18% of Iowa's demand growth since 2021 — so the grid benefit for residents is marginal next to the data center load.

Mistral raised €3 billion at a post-money valuation above €21 billion, led by Samsung Electronics with EQT's Scaleup Europe Fund and PSG Equity as co-leads, plus BlackRock, Advent, and the Grand Duchy of Luxembourg joining a16z, Nvidia, and ASML. The money funds 1 GW of European compute by 2030 and a sovereign-AI pitch that now sells in 20 countries — proof that non-American origin has become a priced commercial asset, not just a political talking point. The cap table stays resolutely international, so 'sovereign' here means customer control over model and region, not European-only capital.

NYU's Tristan Buckmaster and Anthropic's Levent Alpöge released three Lean-verified preprints proving finite-time blowup with smooth forcing for the incompressible porous medium equation, the 2D Boussinesq system, and 3D Euler — real results, drafted largely by Claude and Codex under human direction, with Buckmaster calling the first model output the worst mathematical writing he had ever read. He also disclosed a 6 September call where OpenAI staff described an internal model's roughly 100-page forced Navier-Stokes blowup proof, proposed a coordinated announcement, and twice pushed to drop Alpöge from authorship over the Anthropic affiliation; he declined, and says he has not seen the proof and makes no accusation about training data. Machine-generated proofs have reached the edge of a Millennium Prize problem, and the credit and provenance norms around them are being negotiated in public rather than settled.

China's MIIT set a 2026-2030 target of 9,800 exaflops of intelligent computing capacity, up from 2,185 exaflops in June, backed by 3.8 trillion yuan ($532 billion) in cumulative infrastructure investment and clusters running 10,000 or 100,000-plus accelerator cards. The plan calls for adapting that infrastructure to home-grown chips, making the buildout a demand floor for domestic accelerators — and with 52 facilities above 10,000 cards built and capacity up 177% year-on-year, the fourfold jump extends a curve rather than starting one.