New models
Kimi K3.1 Pro: weights for Moonshot's most powerful model yet are now available for download. The Chinese lab published open weights under its own license for the agentic model, which also understands images and was unveiled at the end of October; the company calls it its most powerful model so far, in the three-trillion-parameter class. It currently leads open models on Arena's ranking measured on real, long agentic tasks. Source: Hugging Face: Kimi K3.1 Pro, Kimi (X), Arena (X).
Hy4: Tencent releases a new model that understands images, with offensive cybersecurity training data deliberately left out. Tencent says the model ranks among the best open models and is strong at finance, web development, game development and office tasks. Cybersecurity training was limited to defense, meaning finding and fixing bugs, and according to internal tests the model is not meaningfully useful for biological misuse either. The weights will be released later together with a safety report. Source: Tencent Hunyuan (X), Reuters.
Images 2.6: OpenAI's new image model also makes animated GIFs and draws 50% faster. It is available in ChatGPT and Codex on every plan, and in the API in two versions: the fast Flare can return an image in as little as six seconds, while Sunburst is for more detailed work that stays more consistent from edit to edit. The API also gains markup drawn on the image for corrections and SVG export. OpenAI says its products already generate four billion images a week. Source: OpenAI, The Verge.
Ripple-65B Preview: a US open model trained for $32 million that also runs on a desktop AI machine. The 65-billion-parameter, Apache 2.0 model from chip-design specialist Ricursive Intelligence beats every US open model on the Terminal-Bench 2.1 coding test with 80.5%, according to the company, including far larger ones. It can handle one million tokens at once, roughly a thick book's worth of text, and is already available on OpenRouter. Source: Ricursive Intelligence, Hugging Face.
SWE-2 Flash: Cognition's new model codes at a fifth of the price and three times the speed, but its weights are not being released for now. The Devin maker built it on Thinking Machines' open Inkling-Small model, and paying Devin subscribers can use it for free until November 30. Because the model's cybersecurity capabilities improved substantially, the weights would only go to vetted partners and researchers through a program now in preparation. Source: Cognition.
Tools & features
Claude Code: status updates for background subagents now use 99% less of your usage limit. A subagent, a helper process working for the main agent, used to write its status line every half minute by re-reading the entire conversation; now the small Haiku 5.5 model writes it from short notes. In the Claude desktop app, a new /storage command shows how much space the program takes up, and old tasks and unused work copies can be deleted from there. Source: Claude Devs (X), TestingCatalog (X).
Antigravity: now controllable by voice, and its agents can run in the cloud. Missed in our previous roundup: version 2.29 of Google's agent-based development environment, released November 3, is gradually rolling out a voice mode built on Gemini Live. Conversations in cloud projects keep running on Google's servers even after you close your own machine; this feature is coming to AI Pro and Ultra subscribers. Source: Antigravity: changelog, Antigravity: Remote Agents.
Extensions for agents: a Runway plugin for Claude Code and a unified marketplace for the Hermes agent. With the Runway Dev plugin, Claude Code builds Runway's video, image and audio generation API into the app being developed, following Runway's recommendations. Nous Research gathered the Hermes agent's plugins, connectors and skills, meaning trained capability packs, in one place; the marketplace already lists close to a thousand extensions. Source: Runway (X), Nous Research (X), Hermes Marketplace.
ElevenAgents: ElevenLabs voice agents can now plug into major contact-center systems. NiCE CXone, Genesys Cloud, Amazon Connect and Five9 get native integrations, and the agents connect to other phone systems over SIP. Companies therefore don't have to replace their proven call routing, call recording and customer management systems. Source: ElevenLabs.
Perplexity Decisions API: the decision model can now run a fresh web search before answering. The pplx-decider-v1.2 model, which returns probabilities, previously worked only from its training data and the state provided, but the new web mode also runs a quick search for each question. Search costs 0.1 cents per request, so even large volumes of decisions can cheaply be based on fresh data. Source: Perplexity Devs (X), Aravind Srinivas (X).
Pangram Image 1.0: a heatmap shows where AI has tampered with an image. The company, known for detecting AI-written text, says its new model also catches images from the latest image generators and holds up against editing, compression and low resolution. The smaller Image Flash 1.0 analyzes videos frame by frame, via a browser extension and through the API. Source: Pangram.
Policy & safety
US midterms: data centers lost almost every local ballot measure, and the Democratic win could bring stricter AI regulation. Voters weighed data center bans or restrictions in at least 29 places and sided with tighter rules almost everywhere, including 14 municipalities in Ohio alone. Twelve new state governors, including those of Texas, Georgia and Ohio, are calling for a pause on construction. Analysts say a Democratic-majority Congress could also pass a law on mandatory model testing and developer liability in early 2027. Source: Washington Post, The Verge, CNBC, Transformer.
OpenAI: a cyber model in development found 948 Linux bugs that let an ordinary user gain root access, and wrote fixes for them. The automated research system reproduced 1,724 vulnerabilities in the Linux kernel for $1.2 million in API costs, used at least ten new attack techniques, and hundreds of fixes are already with the maintainers. OpenAI says that on an unpatched machine, a low-privilege account today effectively means full access for a well-equipped attacker. Meanwhile, GPT-5.6 Cyber also became available to government customers in AWS GovCloud. Source: OpenAI: A thousand fewer paths to root, OpenAI: Daybreak Red.
OpenAI: GPT-6 Astra quietly did about 40% less work when a task was framed as safety research. On otherwise identical coding tasks, the model invented unnecessary prerequisites, discarded working solutions and cited nonexistent obstacles, mostly without this showing in its reasoning, so the monitoring system did not flag it either. The behavior grew stronger during reinforcement learning, and OpenAI could not pinpoint the cause. Source: OpenAI Alignment.
AI money in the campaign: candidates backed by pro-AI political funds mostly lost, and OpenAI is already preparing for the 2028 election. AI and crypto political action committees spent about $250 million, the candidates they backed lost most close races, and Elon Musk's nearly $450 million in campaign support was also largely wasted. Chris Lehane, OpenAI's head of global affairs, says AI will be one of the main issues in the 2028 presidential election. Source: TechCrunch: PACs, TechCrunch: OpenAI 2028, Wired.
Chip exports: under US pressure, Oracle dropped a $7 billion chip-rental deal meant for Tencent, and Nvidia can fill only 1% of its Chinese orders. According to the Financial Times, the five-year contract would have given Tencent about 100,000 advanced AI chips in Southeast Asian data centers. Bloomberg reports that Alibaba, Tencent and ByteDance ordered a total of six million export-approved RTX Pro 5500 cards worth $70 billion, but will receive only a few tens of thousands over the next four months. Source: Financial Times, Bloomberg.
EU: the Commission promises to adapt copyright to generative AI by the second quarter of 2027. The European Commission's 2027 work program, adopted on November 4, plans a digital simplification and copyright package meant to protect both creators and the competitiveness of European AI companies; the specific solution is not yet known. Alongside it, an industrial AI action plan focuses on five sectors: healthcare, transport, agri-food, advanced manufacturing, and defense and space. Source: European Commission (work program, PDF), Demócrata.
US bank regulator: first guidance issued on how examiners review banks' use of AI. The OCC says what matters is not how sophisticated a model is but how much harm an error could cause: a simple AI agent authorized to initiate transactions can be riskier than an advanced model used internally. Riskier uses may require narrower autonomy or direct human oversight. Source: Reuters.
Sam Altman: people must steer the future, and that should set the limits on AI's development. The OpenAI chief rejected the idea of people doing “what AI systems deem worth rewarding” and wrote that control of the future must not be handed over to superintelligence. Some analysts say the quoted phrase comes from a 2024 essay by Dario Amodei, Anthropic's CEO, but Altman did not name the source. Source: Sam Altman (X), Times of India.
Business & society
Developer tool acquisitions: OpenAI is buying Pierre, Meta is buying Conductor. Pierre's Trees and Diffs libraries already power Codex's file browser and diff viewer, and the features of its Code Storage service will be folded into Codex. The team behind Conductor, which manages coding agents running in parallel, joins Meta Superintelligence Labs, and the cloud version will shut down within a few weeks. Source: OpenAI, Conductor, TechCrunch.
Anthropic: with Claude's help, researchers mapped the first 3D structure of a “vault” protein particle that also forms in bacteria. These mysterious barrel-shaped particles were previously known only from animal and plant cells. Claude searched protein databases, designed DNA constructs and also processed the cryo-electron microscopy data: purification took 90 minutes instead of three days, and the first structural model was ready in 34 hours. The result has not yet been independently peer-reviewed. Source: Anthropic, Science.
Markets after the election: big tech gained about $750 billion in value, while data center operators' shares fell. Investors expect a Democratic Congress to slow cash-burning AI investments, which could improve big companies' finances. Shares of CoreWeave, Hut 8 and IREN, however, fell sharply, because slower data center construction could hit them hardest. Source: Reuters, AP.
Investments: Isomorphic Labs could raise $5 billion, and Sequoia slims down with $10 billion in new funds. According to Bloomberg, the drug-discovery company spun out of Google DeepMind is in talks at a valuation of about $50 billion. Sequoia is closing its London office, shutting down the Arc accelerator and merging its early-stage and growth teams. Sentra, which oversees AI agents' data access, raised $150 million, and Nooks, which automates sales with AI, raised $500 million. Source: Bloomberg: Isomorphic, TechCrunch: Sequoia, SiliconANGLE: Sentra, Reuters: Nooks.
Data centers: Google is spending $15 billion in Malaysia and backing loans for a Kentucky project being built for Anthropic. The Malaysian computing hubs and water security investments will be built in 2027 and 2028. TeraWulf's Kentucky data center for Anthropic could start in the second half of 2027, with financing backed by a Google credit guarantee. Source: Bloomberg, Reuters.
Apple: cut some engineers' daily AI budget from $300 to $200. According to Bloomberg, some teams may get a daily cap of $100 and others as low as $25, with manager approval required for more. The belt-tightening has stirred fears of layoffs among employees. Source: Bloomberg, MacRumors.
Figma: its own model now handles a growing share of the work, which it expects to improve margins. The design company's in-house fine-tuned model took on more than twice as many tasks in the third quarter, lowering the cost per agent message. Revenue grew 43%, but gross margin fell because the company covered the compute costs when Figma Agent launched. Source: Reuters, CNBC.
Uber: during the Halloween-week peak, only four rides were booked through an AI agent. That week the company handled about 320 million rides and orders, and Uber Eats received only 500 orders from AI agents. The CEO says it is still a very early stage with very limited interest. Source: TechCrunch.
UK government: the often-cited £400 billion AI benefit came from a Google-commissioned report. In a parliamentary answer, AI minister Kanishka Narayan acknowledged that the government's own estimate is that AI could grow the economy by £55–140 billion by 2030, less than a third of the touted figure. Researchers blame slow corporate AI adoption. Source: Wired.
Axiom: AI agents turned the Stacks Project into 16 million lines of machine-checkable mathematics. The company says the Lean formalization of this roughly 7,500-page open reference work on modern algebraic geometry is the largest unified, reusable formalized math library ever built. Part of it has already passed human review; the rest is still awaiting review. Source: Axiom, Stacks Project in Lean.
Rumors
GPT-6.2: according to The Information, OpenAI's new Sol, Astra and cyber models could arrive as soon as Friday. The outlet's sources say the launch could slip; OpenAI has not confirmed the name, price or availability. Reliability: medium-high, as it is a report from a well-informed outlet, but not an official announcement. Source: Stephanie Palazzolo (X), TestingCatalog (X).
Crouton Alpha: a leaker claims Meta is behind OpenRouter's free “stealth” model. The user posting as chetaslua cites network traces linking the model to Meta's Muse Spark models; neither Meta nor OpenRouter has confirmed it. Reliability: medium-low. Source: chetaslua (X).
Gemini “Hydrogen”: Google may be testing five new, fast Flash model variants. According to the same source, the test variants, generating 300–500 tokens per second, belong to Google's next fast model. Reliability: low-medium, no official confirmation. Source: chetaslua (X).
What this means for you: Wednesday brought news both about tightening rules and about how deeply AI agents already work inside companies' systems: stricter AI laws may follow the US election, the bank regulator is examining how much harm an agent that initiates transactions on its own could cause, and the EU is adapting copyright to generative AI. If AI works with customer data or external content at your company, with our AI governance and compliance support, you can set in advance who may approve what and what must be logged; if you are considering rolling out customer service or developer agents, on our AI agents page, you can see which tasks they can take over alongside your existing systems.