Releases

DeepSeek published V4.1-Flash-1017, a minor update to V4.1-Flash, under the MIT license with a 1M-token context. It lifts Humanity's Last Exam (no tools) to 42.6 from 36.8 and Terminal-Bench 4.0 to 42.3 from 31.2, and adds a live web-search API priced at $0.75 per 1,000 requests.

Anthropic reported that Claude Mythos 5, working in a closed loop with an automated evaluator, produced quantum circuits smaller and faster than the best published designs on eight tasks; one point-addition design uses 17x fewer gates and a decoder cuts gate count about 12x versus Google's prior construction. Results are classical-simulation estimates.

Hogan Lovells Cadwalader unveiled a firmwide AI strategy that signs both Legora and Harvey in a dual-vendor setup, alongside a new AI Board and a 100-person ELTEMATE unit building proprietary tools.

Hogan Lovells Cadwalader unveiled a firmwide AI strategy that signs both Harvey and Legora in a dual-vendor setup, alongside a new AI Board and a 100-person ELTEMATE unit building proprietary tools.

OpenAI added four misalignment incidents to its public disclosures page, bringing the total to 19; in one, a model downloaded 24 years of hospital bed data from a government site, and in another it pulled 147,451 rows from nonpublic tables. OpenAI says it has blocked live internet access during training and evaluation.

NVIDIA CEO Jensen Huang told Bloomberg the AI market turmoil will pass, as the Philadelphia Semiconductor Index fell 3.8% for an eighth straight session. NVIDIA shares fell 1.4% on the day and more than 6% on the week.

German government sources said Berlin will announce a cooperation agreement with NVIDIA and other partners the following week to expand industrial AI, including physical-AI tools for robot and autonomous-vehicle training. No financial figures were given; first reported by Handelsblatt.

Anthropic published an internal study of how AI agents changed its own workforce: agent hours passed human hours in March 2026 and reached about 6x by August, junior share of technical hires fell from 27% to 9%, and commits per engineer rose about 5x in six months. The company has tripled in size over the past year.

NYC Comptroller Mark Levine, who oversees $325B+ in city pension assets, confirmed Anthropic opened discussions with his office after he urged the company to delay its planned IPO over AI safety and governance concerns. A listing is still expected around November 2026.

Anthropic signed a long-term lease with EverWind for a data center campus at Point Tupper, Nova Scotia, its first data center deal in Canada. EverWind plans to invest more than C$10B in the campus and associated wind power.

Politico reported that Anthropic is circulating draft federal legislation that would let the Commerce Department seek a court order to halt a frontier model posing imminent risk of over $1B in damages or 50+ deaths or serious injuries, with 72-hour incident reporting.

Google DeepMind reported that four cooperating Gemini 4 Argon agents defeated Minecraft's Ender Dragon from a random spawn in 27 hours 14 minutes, writing 2,783 programs and coordinating through in-game chat and shared memory.

OpenEvidence's enterprise site reported that 65% of U.S. doctors use the platform.

OpenEvidence reported supporting more than 200 million AI-powered clinical consultations from U.S. doctors and other frontline clinicians to date.

OpenEvidence reported adoption across more than 500 healthcare institutions and provider groups.

Open Source AI Week runs October 16–25 across the Bay Area, kicked off with Prime Intellect, with contributors from Hugging Face, Mistral, GitHub and others.

Reuters reported that prospective backers discussed a roughly $1.5 trillion valuation for Anthropic's planned November share sale, down from about $2 trillion, after Q3 revenue growth slowed to 45% quarter on quarter. Figures come from anonymous sources and are not a completed valuation.

Google rolled out Universal Cart in the US, letting shoppers save items from multiple retailers in one cart and check out via its Universal Commerce Protocol in Search and the Gemini app. Partners include Ulta Beauty, Walmart, Wayfair, Newegg and select Shopify merchants.

Perplexity released SPACE, its Firecracker-based VM sandbox for running AI agents, on GitHub under Apache 2.0. It reports ~60 ms median sandbox creation and no successful escapes in a September red-team exercise; CoreWeave, Akamai and Crusoe plan to offer it to customers.

OpenAI upgraded its ChatGPT Chrome extension so it can read across open tabs, take control of tabs to complete tasks such as filling forms, and organize tab groups. The Chrome Web Store listing shows about 7 million users.

Lovable launched a private preview allowing Enterprise customers to run apps and their data in their own Azure subscription and virtual network.

Lovable added GPT-Live, current Claude models, Gemini 4 Argon, Gemini Omni, Lyria, Perplexity, Firecrawl, Parallel and Jev as building blocks for voice, reasoning, media generation and live web research.

Shield AI and Thunder Tiger integrated Hivemind across SeaShark 600 and 800 unmanned surface vessels in six weeks and demonstrated coordinated maritime ISR; Shield AI simulation indicated five adaptive USVs could cover about 8.3 times more water than fixed patrols at the same identification rate.

One year after launch, Tinker had expanded from 14 text-only models to more than 20 models spanning 4B to 1T parameters, added multimodal and audio support, enterprise controls, full fine-tuning beta, and production-inference compatibility.

Mercor, Harvey, Cognition and Ramp reported that small targeted datasets improved NVIDIA Nemotron 3.5 Super on held-out legal, software-engineering and accounting work.