AI Weekly · Week 45, 2026 · Nov 2 – Nov 8
ICREAT AI NEWS
This week saw dense model-side updates: Google DeepMind's Gemini Omni 1.2 Flash extended native video generation to 15 seconds and cut prices by 30%, and Tencent's Hy4 added visual understanding and computer-use for the first time. On the product side, Apple opened Siri AI to all developers, and Cursor can directly generate and deploy online apps. On the industry side, Anthropic's quarterly revenue grew twelvefold but it posted a net loss of $23.3 billion, and its $1.5 trillion IPO hit a snag over valuation concerns.
GPT-6 Astra directly rewrites other models' weights through weight-space programming
The Decoder reports that GPT-6 Astra uses a method researchers call Weight-Space Programming, directly reading and rewriting the raw weights of other models without going through the conventional training process of step-by-step mathematical optimization.
The DecoderOriginalModel Releases/Updates
Section overviewNew models are generally expanding toward longer videos, stronger agentic operation, and security-specific directions, while the gap between open-weight and closed-source models continues to narrow.
Google DeepMindFirst party11.06Google DeepMind releases Gemini Omni 1.2 Flash, supporting 15-second videos with a 30% price cut
Google DeepMind releases Gemini Omni 1.2 Flash, raising native video generation length from 10 seconds in Omni 1.1 to 15 seconds, and supporting per-second specified durations from 3 to 15 seconds. The model adds multi-turn conversational editing and stronger reference image consistency; 720p video is priced at $0.07 per second, 30% lower than Omni 1.1, and is now generally available through the Gemini API and Gemini for business.
OpenAI releases GPT-6 Cyber cybersecurity model
OpenAI releases the cybersecurity model GPT-6 Cyber, used to discover exploitable vulnerabilities and develop remediation plans. The model is trained and instructed to act only within authorized scope.
Thinking Machines releases Inkling 2 Small, weights pending release
Thinking Machines releases Inkling 2 Small, which scores 31 on the Artificial Analysis Intelligence Index v5.0.1, 18 points higher than Inkling Small and on par with Gemini 3.8 Flash (high).
Tencent releases Hy4 open-source large model, first to support visual understanding and computer-use
Tencent releases its next-generation large language model Hy4, which it says ranks among the world's leading open-source models, with model weights and system card to be released in the coming days. Hy4 introduces visual understanding and computer-use capabilities for the first time, can operate command lines and graphical interfaces, and shows notable improvements in instruction following, long context, and long-horizon Agent tasks.
Product Releases/Updates
Section overviewDeveloper entry points continue to open up, Siri AI opens to all apps, and inference and deployment tools are also accelerating adaptation to new models.
Apple opens Siri AI to all developers and lets developers teach it skills
Apple opens the Siri AI developer preview in the first beta of iOS 27.3; any app can connect its own content and actions to Siri AI through App Intents, no longer limited to preset categories.
vLLM BlogFirst party11.04vLLM supports Kimi K3.1 Pro on day one, covering NVIDIA, AMD, TPU, and Zhenwu PPU
vLLM announces support completed on the day Kimi K3.1 Pro was released, covering NVIDIA GPUs, AMD GPUs, Google TPUs, and Alibaba T-Head Zhenwu PPU. K3.1 Pro is Moonshot's 2.8T-parameter (104B activated) multimodal MoE open-weight model, adding DSA sparse attention to all 24 full-attention layers, with a 1M context window.
Cursor ChangelogFirst party11.05Cursor launches app-building feature that can directly generate and deploy shareable online apps
Cursor launches an app-building feature; after users describe their requirements in the cloud agent, Cursor builds the web app, creates a Postgres database as needed, configures Cursor SSO authentication, and completes hosted deployment. When publishing, users simply choose a domain to get a shareable link.
Together AI BlogFirst party11.04Together AI releases Together Compiler, using agents to continuously compile and optimize AI inference
Together AI releases Together Compiler (TKCC), which uses high-throughput agents to continuously generate optimizations for inference engines, achieving up to 2.6x TPS improvement on frontier open-source models in under 24 hours.
Industry news
Section overviewCapital and policy are tightening simultaneously: China plans to impose export controls on frontier AI technologies, and companies are beginning to use self-hosted open-source models to lower coding costs.
The Decoder11.03Anthropic quarterly revenue grows twelvefold, but net loss reaches $23.3 billion
Anthropic's revenue this quarter is nearly $17 billion, twelve times that of a year ago, but operating costs of $17.5 billion exceed revenue, resulting in a net loss of $23.3 billion, compared with $12.1 billion in the same period last year. The company attributes the cost increase to a $5 billion short-term compute lock-in and $2 billion in employee donation matching; the updated prospectus shows compute and data center lease commitments of $895 billion. The IPO is expected to raise about $100 billion, of which up to $3.5 billion comes from Japanese investors.
China releases its first draft frontier AI export control rules: Ministry of Commerce Announcement No. 48 of 2026
China's Ministry of Commerce and Ministry of Science and Technology issued Announcement No. 48 on October 16, 2026, proposing to add 40 technologies to the Catalog of Technologies Prohibited or Restricted from Export in China, 23 of which belong to artificial intelligence, while the already-effective transitional pre-export registration requirement is implemented simultaneously.
The Decoder11.06Anthropic's $1.5 trillion IPO hits a snag as investors question the valuation
Anthropic's IPO has hit a snag as investors object to the $1.5 trillion valuation, with some potential investors doubting whether the roadshow originally scheduled to start next week can proceed as planned. The company had targeted a valuation of up to $2 trillion in September; now some investors believe $1.5 trillion is still too high, and some have even suggested only $1 trillion.
Sophos discloses that the Aur0ra ransomware gang uses AI models such as Cursor to disguise attacks as red-team tests and bypass guardrails
Sophos's threat research unit discloses that the Aur0ra ransomware operated by the GOLD OGRE group makes extensive use of AI during intrusions, bypassing model guardrails by packaging attacks as authorized red-team tests or diagnostic tests.
Democratic midterm election victory raises AI investment concerns, global stock markets plunge
After the Democrats' landslide victory in the U.S. midterm elections, markets worried that the AI investment boom would be undermined, and global stock markets fell sharply. More than $3.2 trillion in U.S. stock market value was wiped out, the Nasdaq fell 5.5%, Nvidia dropped more than 9%, SanDisk fell 22%, and Micron and AMD each fell 17%. Asia-Pacific markets also fell in tandem, with the Nikkei down nearly 6%, South Korea's KOSPI down 6.7%, and Taiwan's TAIEX down 4.5%.
Revolut replaces Claude and Codex with self-hosted GLM, cutting AI coding bills by about 90%
Revolut has switched the default coding tools of about 2,000 software engineers from Anthropic's Claude Code and OpenAI's Codex to the open-weight model GLM from Chinese lab Z.ai, self-hosted by the company, reducing bills to about one-tenth of the original.
The Decoder11.04OpenAI says GPT-6 Astra designed a hardware device from scratch in under a week
OpenAI says its hardware team used GPT-6 Astra to design a device that lets users talk to the Codex coding agent, completing it as a side project in under a week. Astra handled everything from enclosure design, PCB layout including component selection and routing, firmware, and user interface; a contract manufacturer produced a small batch of prototypes in under a month, but the design went through multiple revisions because manufacturing constraints, tolerances, and assembly details were not fully considered. OpenAI plans to feed this data back into subsequent models.
Research Papers
Section overviewMultiple evaluations show that agents still have clear shortcomings in real scientific research and cyber offense/defense tasks, and measuring the pace of model capability improvement is itself full of uncertainty.
The Decoder11.07Nine AI agents take on open research questions from U.S. national laboratories; none of 135 submissions meets expert standards
A new paper had nine AI agents, including Claude Opus 4.8, GPT-6 Astra, and DeepSeek V4 Pro, conduct follow-up research on 15 open questions from researchers at U.S. national laboratories; none of the 135 submissions simultaneously met the three standards of honesty, completeness, and expert quality.
METR BlogFirst party11.05METR independently evaluates Anthropic's AI R&D acceleration in October 2026
METR conducted a high-permission evaluation of Anthropic from August 12 to October 29, 2026, concluding that as of October 2026, Anthropic's rate of model capability improvement on well-defined tasks was about twice that of March 2024 to October 2025, but with very high uncertainty, with the true value possibly ranging from 1.3x to 3x.
The Decoder11.07Vals AI benchmark shows the gap between open-source and closed-source models in cyber offense capability continues to narrow
Vals AI's CyberOffense benchmark has AI agents attack a simulated corporate network from scratch on an isolated Kali Linux machine; GPT-6 Sol Daybreak in OpenAI's restricted program completed 80.8% of objectives, while the open-weight model Deepseek V4.1 Flash 1017 reached 72.8%, at a cost of about $3 per run instead of about $22.
Tips and Opinions
The cost of AI breaching corporate networks fell 99% in five months, and security budgets are shifting to a compute competition
Tomasz Tunguz points out that the cost of AI breaching corporate networks fell 99% in five months: in April, Claude Mythos Preview and GPT-5.5 autonomously completed a 32-step intrusion that would take a human expert 20 hours, with BIS estimating the cost at $5,000 to $10,000; by August, Zhipu's open-source GLM-5.3 had brought it down to $90.