# AI Headline Hub > AI Headline Hub covers AI models, products, business, policy, and funding, with links to primary sources. AI Headline Hub publishes reported stories and links primary sources. Search crawlers and AI crawlers are explicitly allowed in robots.txt, including GPTBot, ClaudeBot, CCBot, Google-Extended, Bytespider, and PerplexityBot. Pages do not use noai or noimageai tags. That access policy does not assign copyright. Full story text is on each article page and in the RSS and Atom feeds. ## Stories - [Trump and Xi quietly agreed on one AI rule, and it is about who gets to launch a military strike](https://aiheadlinehub.com/2026/11/21/trump-xi-agree-humans-must-stay-in-military-ai-decisions-bessent/): Scott Bessent says Trump and Xi agreed in Shenzhen that humans should always be involved in military decisions made with AI. China’s readout says AI must stay “under human control”. - [OpenAI says rivals tried to copy its models' hidden reasoning, and its new defences cut one attack's success from 54% to 0.2%](https://aiheadlinehub.com/2026/11/21/openai-adversarial-distillation-defenses-reasoning-extraction-54-to-0-2-percent/): OpenAI says layered safeguards in GPT-6.2 Astra cut one reasoning-extraction attack’s success rate from 54.0% to 0.2%, and that it shares indicators of adversarial distillation with other labs. - [200,000 people quietly installed Meta’s Muse agent in Chrome before Meta even announced it](https://aiheadlinehub.com/2026/11/20/meta-muse-chrome-extension-200000-users-sidebar-agent-email/): Meta's Alexandr Wang announced a Muse Chrome extension that runs the agent from a sidebar using your logged-in sites. The store listing shows 200,000 users. - [OpenAI’s new voice model can listen while it talks, and now it costs less to keep it waiting](https://aiheadlinehub.com/2026/11/20/openai-gpt-live-1-1-voice-api-standby-price-cut-accents/): OpenAI released GPT-Live-1.1 in its API. The full-duplex voice model has steadier accents across languages and a new standby rate of $0.02 a minute when a WebRTC session is paused, against $0.05 a minute when active. - [Google just handed developers its world model: Genie 4 can now build playable 720p worlds with sound through the Gemini API](https://aiheadlinehub.com/2026/11/20/genie-4-gemini-api-developers-build-playable-worlds-720p-native-audio/): Genie 4 is now available to developers through the Gemini API and Google AI Studio, a day after launch. It generates explorable 720p worlds at 24fps with native audio from a text prompt, an image or both. - [Perplexity's agent now writes its own skills by watching how you work, and it has already written 80,000 of them](https://aiheadlinehub.com/2026/11/19/perplexity-brain-writes-its-own-agent-skills-80000-personal-skills/): Perplexity says Brain, the personalisation system in its Computer agent, now creates and updates reusable skills from your past sessions. It reports 80,000+ personal skills written for 20,000+ users since August. - [OpenAI's agents now finish almost half of its 2 to 4 hour research tasks alone, and the median researcher uses $2,400 of AI a day](https://aiheadlinehub.com/2026/11/19/openai-agents-finish-half-of-research-tasks-researchers-spend-2400-a-day/): OpenAI's new research acceleration report says agent success on 2 to 4 hour research tasks rose from 23% to 46% since March, while median daily agent use per researcher hit $2,400 at list prices. - [OpenAI says its models found a year's worth of gains in AMD's flagship GPU design in under five months](https://aiheadlinehub.com/2026/11/19/openai-models-amd-gpu-14-percent-ppa-jalapeno-chip-design/): OpenAI says its models found up to 14% PPA gains on AMD's flagship GPU IP and cut Jalapeño's vector unit area by about 13%, with some changes already sent to manufacturing. - [Google's Gemini 4 Neon lands between Carbon and Hydrogen, at $3 per million input tokens](https://aiheadlinehub.com/2026/11/19/google-gemini-4-neon-launch-price-benchmarks/): Sundar Pichai announced the model on X. Google says Neon ranks third on the Artificial Analysis Intelligence Index and costs 32% less per task than Carbon at its default setting. - [OpenAI lets outside safety assessors inside, but critics say it is far from 'employee-like access'](https://aiheadlinehub.com/2026/11/19/openai-embedded-safety-assessors-critics-access/): Staff from four safety groups will work as embedded assessors. Researchers say their access is narrow, with no internal Slack and limits on which OpenAI employees they can talk to. - [OpenAI ships GPT-6.2 Astra and Sol, and Sam Altman singles out Sol's Ultrafast mode](https://aiheadlinehub.com/2026/11/18/openai-gpt-6-2-astra-sol-launch-altman-ultrafast/): The two models are live in ChatGPT Work, Codex and the API. Astra's API price drops 20%, Sol keeps GPT-6.1 Sol's rates, and Ultrafast reaches Pro 200 subscribers for the first time, starting with Sol. - [Microsoft will license MAI model weights so companies can run and customise them in their own Azure tenants](https://aiheadlinehub.com/2026/11/18/microsoft-mai-licensed-weights-azure-ignite-2026/): Announced at Ignite, MAI Licensed Weights is coming soon through Microsoft's Frontier Company engagements. Microsoft has not published prices or licence terms. - [Isomorphic Labs raises $5 billion to push its AI-designed drugs towards clinical trials](https://aiheadlinehub.com/2026/11/17/isomorphic-labs-5-billion-series-c-greenoaks-sequoia/): Greenoaks and Sequoia Capital co-led the Series C, with a16z and Alphabet also investing. Demis Hassabis says the money will fund biology models, disease research and the first programmes heading to the clinic. - [Hugging Face hands language-model routing for its Inference Providers to OpenRouter](https://aiheadlinehub.com/2026/11/17/hugging-face-inference-providers-switch-to-openrouter-routing/): Chat-model requests made with a Hugging Face token now go through OpenRouter, which picks a provider on price, speed and reliability and retries on failure. Own provider keys no longer work for these models. - [OpenAI opens a science blog with a physicist's account of using GPT-6 Astra on the Standard Model](https://aiheadlinehub.com/2026/11/16/openai-science-blog-gpt-6-astra-standard-model-lattice/): UC Santa Barbara physicist Nathaniel Craig says GPT-6 Astra helped turn his idea for a lattice definition of the Standard Model into a detailed proof that still needs expert scrutiny. - [Apple launches a study app to test how people use AI for health questions](https://aiheadlinehub.com/2026/11/16/apple-openevidence-health-ai-study-app/): Apple and OpenEvidence are inviting US iPhone users to a study of up to 200,000 people over more than a year. Answers draw on clinical sources, and Apple says the data will be anonymised and not used to train AI models. - [Perplexity CEO says Teammate agents now add 11 times more work hours a week than in September](https://aiheadlinehub.com/2026/11/15/perplexity-teammate-11x-agent-hours-srinivas/): Aravind Srinivas says agent hours from Perplexity Teammate are now 11 times the September level, up from 7 times a week earlier. The figures are his own claims. - [Gemini can now read Google Wallet spending data for US users](https://aiheadlinehub.com/2026/11/15/gemini-google-wallet-personal-intelligence-us/): Google has added Wallet to Gemini's Personal Intelligence connections in the US, with optional bank and card links through Plaid. Gemini can chart spending and flag subscriptions but cannot move money. - [Anthropic publishes its playbook for building AI safety evaluations](https://aiheadlinehub.com/2026/11/15/anthropic-safety-evaluation-design-recommendations/): A 13 November research post sets out how Anthropic designs, sandboxes, grades and reports safety evals, from multi-turn simulated users to checking LLM graders against human experts. - [OpenAI launches ChatGPT Wallet so dots can pay for purchases without seeing card numbers](https://aiheadlinehub.com/2026/11/14/chatgpt-wallet-dots-amex-mastercard-visa/): ChatGPT Wallet, built with American Express, Mastercard and Visa, uses network payment tokens and a security check on every purchase. It starts with US Pro subscribers using dots. - [llama.cpp and ggml reach version 1.0.0, pruning deprecated APIs](https://aiheadlinehub.com/2026/11/14/llama-cpp-ggml-v1-0-0-release/): The open source local inference projects shipped 1.0.0 releases on 13 November, removing deprecated public APIs, adding FP8 support and hardening llama-server. - [Google launches Gemini 4 Hydrogen, a faster everyday model that it says beats Argon on one index](https://aiheadlinehub.com/2026/11/13/gemini-4-hydrogen-fast-model-launch/): Google's new Gemini 4 Hydrogen costs $1.50 per million input tokens, has four thinking levels and is rolling out across AI Studio, the Gemini app and Search. - [DeepL's Aure voice model carries tone and pace into live translation in eight languages](https://aiheadlinehub.com/2026/11/13/deepl-aure-voice-translation-model/): DeepL says Aure, now inside DeepL Voice and its Voice API, adapts translated speech as a speaker's emotion, emphasis and pace change, with no setup needed. - [Google launches Gemini Decisions API, scoring up to 1,000 questions in one call](https://aiheadlinehub.com/2026/11/12/gemini-decisions-api-gemini-decision-1/): Google's new Gemini Decision 1 model answers yes/no, choice and rubric questions in one pass, at $0.10 per million input tokens and no output charge. - [Google DeepMind releases AlphaGenome 2, predicting DNA variant effects cell by cell](https://aiheadlinehub.com/2026/11/12/alphagenome-2-cell-type-dna-variant-model/): AlphaGenome 2 predicts how DNA variants act in specific cell types and finds nearly twice as many expression-changing variants, DeepMind says. - [Alibaba opens a two-week Qwen4-Max preview on its Token Plan](https://aiheadlinehub.com/2026/11/12/qwen4-max-preview-token-plan/): Qwen4-Max-Preview, Alibaba's next flagship multimodal model, is available on Token Plan only from 11 to 25 November, after which the preview is retired. - [Databricks turns Genie into an always-on agent and makes Genie Ontology generally available](https://aiheadlinehub.com/2026/11/12/databricks-always-on-genie-ontology-ga/): Databricks' new Genie runs in the background, delegates to subagents and flags decisions in an inbox; Genie Ontology is now generally available. - [Runway releases Praxis-1 robot model with open weights under Apache 2.0](https://aiheadlinehub.com/2026/11/12/runway-praxis-1-open-weights-robotics/): Runway's 27B-parameter Praxis-1, which predicts video and robot actions together, is now downloadable for commercial use under Apache 2.0. - [MiniMax publishes H3-Image weights, a six-step model for 16-megapixel images](https://aiheadlinehub.com/2026/11/12/minimax-h3-image-open-weights/): MiniMax's 33.1B-parameter H3-Image generates and edits images up to 4096 by 4096 in six steps; its community licence restricts some territories. - [NVIDIA contributes OpenShell to the CNCF to govern autonomous AI agents](https://aiheadlinehub.com/2026/11/11/nvidia-openshell-cncf-sandbox/): On 10 November, NVIDIA said it is contributing OpenShell, an open-source runtime for securing and governing autonomous agents, to the Cloud Native Computing Foundation as a CNCF Sandbox project. - [AWS says autonomous AI agents wrote every line of its new local AWS Emulator](https://aiheadlinehub.com/2026/11/11/aws-emulator-built-by-autonomous-agents/): AWS said on 11 November that AWS Emulator, a local emulator of AWS services, was written entirely by autonomous agents and released on GitHub under a CC0 licence. - [DeepMind's AlphaProof Code formally verifies the Signal protocol's code](https://aiheadlinehub.com/2026/11/11/alphaproof-code-verifies-signal-protocol/): Google DeepMind says its AlphaProof Code agents proved key security properties of Signal's cryptographic implementation in four weeks, a first at this depth. - [Google makes Gemini 4 Carbon available to developers, enterprises and AI Ultra users](https://aiheadlinehub.com/2026/11/11/gemini-4-carbon-now-available/): Gemini 4 Carbon, Google's cyber-focused frontier model, is now in the Gemini API, Gemini for business, Antigravity, the Gemini app and AI Mode, at Argon's introductory price until 30 November. - [What Is a Token in Large Language Models?](https://aiheadlinehub.com/2026/11/11/what-is-a-token-in-llms/): Tokens are the chunks of text a language model reads and writes. How tokenisers split text, why token counts differ from word counts, and how tokens drive limits and cost. - [What Is Prompt Injection?](https://aiheadlinehub.com/2026/11/11/what-is-prompt-injection/): Prompt injection is when text fed to an AI model overrides its intended instructions. Direct and indirect attacks, real examples, and the defences security agencies recommend. - [Anthropic launches Claude Fable 5.2 and Mythos 5.2, one model with two routes to users](https://aiheadlinehub.com/2026/11/10/anthropic-fable-mythos-5-2-models/): Anthropic's new Fable 5.2 is open to all users while Mythos 5.2, the same model with fewer safeguards, stays limited to vetted cyber and life sciences groups. - [Google adds TPU 8tx and TPU 8i Max, its third and fourth AI chip systems this year](https://aiheadlinehub.com/2026/11/10/google-tpu-8tx-8i-max/): Google's TPU 8tx targets confidential inference for regulated users, while the dual chip TPU 8i Max is built for very long context AI workloads. - [MIT's FLEX uses an LLM to help robots read intent into spoken corrections](https://aiheadlinehub.com/2026/11/09/mit-flex-robots-interpret-natural-language/): MIT and NYU researchers built FLEX, a three-step Add, Generalize and Mask pipeline that lets a robot arm generalize from commands like "Don't spill on the laptop" to avoiding all electronics. - [NPCI forms NPCI Tech Solutions, a new subsidiary for AI and blockchain](https://aiheadlinehub.com/2026/11/09/npci-launches-ai-blockchain-technology-subsidiary/): The operator of India's UPI has set up NPCI Tech Solutions Limited, led by former CTO Vishal Kanvaty, to build AI, blockchain and digital public infrastructure technology. - [NVIDIA, Rebellions and SK Telecom team up on Korean sovereign AI inference racks](https://aiheadlinehub.com/2026/11/09/nvidia-rebellions-sk-telecom-sovereign-ai/): Rebellions will connect its R200 NPUs to NVIDIA's platform through NVLink Fusion, and SK Telecom plans to deploy the systems in its AI data center, with R200 systems expected in 2028. - [Perplexity CEO says Teammates automation runs grew 3.7x on day two](https://aiheadlinehub.com/2026/11/08/perplexity-teammates-automation-runs-3-7x/): Aravind Srinivas said on 7 November 2026 that automation runs on Perplexity Teammates, the personal work agents launched on 5 November, grew 3.7 times from launch day to day two. - [Meta introduces Spaces, custom mini-experiences built by describing them in Muse](https://aiheadlinehub.com/2026/11/08/meta-muse-spaces-custom-mini-apps/): Meta launched spaces in its Muse app on 6 November, letting users describe an idea such as a gift tracker or word game and have Muse build it as a reusable custom experience. - [Into the Unknown benchmark finds no AI agent met every expert requirement](https://aiheadlinehub.com/2026/11/08/into-the-unknown-ai-agents-research-benchmark/): None of 135 AI-written follow-up papers met all expert-defined conditions in the Into the Unknown benchmark, and 114 failed its integrity check, researchers report. - [Anthropic IPO Faces Investor Pushback Over $1.5 Trillion Valuation](https://aiheadlinehub.com/2026/11/08/anthropic-ipo-investors-push-back-valuation/): Three major investors are questioning Anthropic's expected $1.5 trillion IPO valuation, with one suggesting a figure closer to $1 trillion as market conditions deteriorate. - [ChatGPT is reportedly building a Search Console-style dashboard for businesses](https://aiheadlinehub.com/2026/11/07/chatgpt-business-hub-mentions-clicks-analytics/): OpenAI appears to be testing a Business Hub tool that would show verified businesses how often ChatGPT mentions them and how many clicks they get. OpenAI has not announced it. - [OpenAI closes about $11 billion at a $1.25 trillion valuation](https://aiheadlinehub.com/2026/11/07/openai-11b-raise-1-25t-valuation/): Reports say OpenAI has completed an initial roughly $11 billion close in a funding round valuing the company at $1.25 trillion, short of earlier $30 billion talk. - [OpenAI backs Frontier Model Forum jailbreak rating framework for cyber threats](https://aiheadlinehub.com/2026/11/07/openai-fmf-jailbreak-rating-framework/): OpenAI said on 6 November 2026 that it partnered with the Frontier Model Forum on a shared framework to rate when bypassing AI safeguards could enable serious cyber harm and to prioritize stronger protections. - [Anthropic details Claude agent memory and offline dreaming](https://aiheadlinehub.com/2026/11/07/anthropic-agent-memory-dreaming/): Anthropic said Claude agent memory uses transcripts, notes, structured forms and a small injected context layer, with offline dreaming to reorganize learning. - [METR: Anthropic’s model progress likely ~2x historical pace, mostly from AI in R&D](https://aiheadlinehub.com/2026/11/06/metr-anthropic-ai-rd-acceleration/): METR’s 5 Nov 2026 assessment says Anthropic is likely improving well-specified task capabilities at about twice the March 2024–October 2025 rate, mostly from AI used in R&D. - [Thinking Machines Lab launches Inkling 2 Small on Tinker](https://aiheadlinehub.com/2026/11/06/thinking-machines-inkling-2-small/): Thinking Machines Lab launched Inkling 2 Small on 5 Nov 2026: a 594B MoE model (17B active) scoring 31 on the Artificial Analysis Intelligence Index, on Tinker now with Apache 2.0 weights planned next week. - [Meta acquires Conductor, an AI coding startup for multi agent workflows](https://aiheadlinehub.com/2026/11/06/meta-acquires-conductor-ai-coding/): Meta has acquired Conductor, whose product helps developers run and coordinate multiple coding agents. Terms were not disclosed. Conductor Cloud shuts on 20 November 2026. - [OpenAI introduces GPT-6 Cyber for aligned security research](https://aiheadlinehub.com/2026/11/06/openai-gpt-6-cyber-security-research/): OpenAI launched GPT-6 Cyber on 5 November 2026, saying the model pairs stronger cybersecurity capabilities with tighter scope alignment for approved Daybreak Red customers. - [Anthropic makes inference hooks generally available for Claude Enterprise](https://aiheadlinehub.com/2026/11/06/anthropic-enterprise-security-hooks-ga/): Anthropic says Claude Enterprise now offers generally available inference hooks alongside new security and administration controls for enterprise customers. - [Google ships Gemini Omni 1.2 Flash with 15-second video and a 30% price cut](https://aiheadlinehub.com/2026/11/05/google-gemini-omni-1-2-flash-15s-price-cut/): Google and Google AI Studio announced Gemini Omni 1.2 Flash on 5 Nov 2026: native clips up to 15 seconds, multi-turn conversational edits, and 720p API pricing at $0.07 per second (30% below Omni 1.1). - [Anthropic: Claude helps map first vault particle from prokaryotic life](https://aiheadlinehub.com/2026/11/05/anthropic-claude-vault-particle-prokaryotes/): Anthropic says Claude helped map the first 3D vault from bacterial/phage life (L05, 96 copies, ~3.2 Å) in a 4 Nov 2026 research post; preprint submitted to bioRxiv. - [TeraWulf secures Google credit backstop for ~$19B Anthropic lease at Kentucky Justified campus](https://aiheadlinehub.com/2026/11/05/terawulf-google-credit-backstop-anthropic-kentucky/): TeraWulf said Google will credit-support project debt for Anthropic's ~$19B Justified leases in Kentucky and signed a Fluor EPC for three buildings, ready from late 2027 through Q2 2028. - [Cat Wu and Austin Lau leave Anthropic for OpenAI](https://aiheadlinehub.com/2026/11/05/cat-wu-austin-lau-leave-anthropic-join-openai/): Claude Code product lead Cat Wu is joining OpenAI to run Core Products PM, and growth marketer Austin Lau says he has also moved to OpenAI, both announcing on X. - [OpenAI launches ChatGPT Images 2.6 with GIFs, reference images and faster API models](https://aiheadlinehub.com/2026/11/05/openai-chatgpt-images-2-6-gifs-flare-sunburst/): OpenAI introduced ChatGPT Images 2.6 with native GIF creation and reference images, plus GPT-Image-2.6 Flare and Sunburst models it says are about 50% faster in the API. - [Factory acquires HumanLayer and will open-source it](https://aiheadlinehub.com/2026/11/05/factory-acquires-humanlayer-open-source/): Coding agent company Factory is acquiring HumanLayer, whose founders Dex Horthy and Kyle Mistele join Factory. HumanLayer will be open-sourced. - [Moonshot AI releases Kimi K3.1 Pro open weights](https://aiheadlinehub.com/2026/11/04/kimi-k3-1-pro-open-weights-nebius/): Moonshot AI has published the weights of Kimi K3.1 Pro, a 2.8-trillion-parameter multimodal model with a one-million-token context window. - [Tencent releases Hy4 with vision and computer use](https://aiheadlinehub.com/2026/11/04/tencent-hy4-vision-computer-use-release/): Tencent has released Hy4, a 770-billion-parameter model that adds vision and computer use, with open weights and a system card promised soon. - [Anthropic's FlashProve aims to prove which AI model answered](https://aiheadlinehub.com/2026/11/04/anthropic-flashprove-provable-inference/): Anthropic has released FlashProve, an open-source prototype that uses zero-knowledge proofs to show a reply came from a committed model without revealing its weights. - [Viral 'Sexbot' Launch on X Is a Parody, Its Own Website Says](https://aiheadlinehub.com/2026/11/03/sexbot-parody-dreamstate/): Dreamstate founder Luis Vale's 'Sexbot' announcement drew hundreds of thousands of views. The product site says there is no robot and calls it a parody. - [Anthropic Adds Smart Reports to Claude Enterprise](https://aiheadlinehub.com/2026/11/03/anthropic-claude-enterprise-smart-reports/): New enterprise reports aim to help administrators assess AI use, cost and efficiency. - [OpenAI Security Warnings Put Research Access Under Scrutiny](https://aiheadlinehub.com/2026/11/03/openai-security-warnings-research-access/): Former employees describe longstanding security concerns as OpenAI emphasizes new protections. - [EU Court Adviser Raises Copyright Questions Over Google Gemini](https://aiheadlinehub.com/2026/11/03/eu-court-adviser-google-gemini-copyright/): An EU court adviser has outlined circumstances in which Gemini answers could infringe publishers’ rights. - [OpenAI's Sottiaux: dots will add GPT-6.1 Terra and more models](https://aiheadlinehub.com/2026/11/03/openai-dots-gpt-6-1-terra/): OpenAI's Tibo Sottiaux says dots, ChatGPT's always-on agents, will run on more models, including GPT-6.1 Terra. No date, price or new plan was given. - [What is a world model? A plain guide to the idea behind Genie](https://aiheadlinehub.com/2026/11/02/what-is-a-world-model/): A world model is an AI system that learns how an environment changes and predicts what comes next. Here is where the idea came from and what it can and cannot do. - [Pika H3 Flex: open weights for a video model with sound](https://aiheadlinehub.com/2026/11/01/pika-h3-flex-open-weight-video-model/): Pika has published open weights for Pika H3 Flex, a video-and-audio model built on MiniMax-H3 that it says generates clips in six sampling steps. - [DeepMind essay says AI coordination, not skill, drives change](https://aiheadlinehub.com/2026/10/31/deepmind-economics-of-intelligence/): A Google DeepMind essay argues that AI agents' ability to coordinate, more than raw competence, decides when AI reshapes the economy. - [OpenAI promises a safety report and tighter teen limits](https://aiheadlinehub.com/2026/10/31/openai-community-safety-report-teens/): OpenAI will publish a twice-yearly Community Safety Report from early 2027 and is adding new limits on violence and weapons content in ChatGPT. - [Google's Suncatcher satellite runs first AI prompt in orbit](https://aiheadlinehub.com/2026/10/31/google-suncatcher-tpu-orbit/): Google says its Project Suncatcher prototype ran a Gemma 4 model on a data-centre-class TPU in space and returned a haiku to the ground. - [AI firms back one robots.txt opt-out for AI training](https://aiheadlinehub.com/2026/10/31/ai-training-policy-token-robots-txt/): OpenAI, Google, Microsoft and Anthropic say they will honour a shared robots.txt policy token so sites can set AI-training preferences once. - [Mistral releases the weights of Mistral Large 4](https://aiheadlinehub.com/2026/10/30/mistral-large-4-open-weights/): Mistral has published open weights for Mistral Large 4, a 1-trillion-parameter model it says leads open models on coding, with a licence condition for model-hosting firms. - [NVIDIA opens a speech model that names who spoke](https://aiheadlinehub.com/2026/10/30/nvidia-nemotron-transcribe/): NVIDIA's Nemotron 3.5 Transcribe is an open-weight model for recorded audio in 40 languages, released with a larger Granary v2 speech dataset. - [Canada plans up to C$1.6 billion stake in Cohere](https://aiheadlinehub.com/2026/10/30/canada-cohere-equity-stake/): Canada's government says it will take an equity stake of up to about C$1.6 billion in Cohere, paid in stages and tied to jobs and computing commitments. - [OpenAI rebuilds its Model Spec around five values](https://aiheadlinehub.com/2026/10/30/openai-model-spec-five-values/): OpenAI's updated Model Spec groups its rules under five values, adds guidance on context and agency, drops most style rules, and calls the model GPT. - [OpenAI details four layers that watch its own agents](https://aiheadlinehub.com/2026/10/30/openai-misalignment-monitoring-layers/): OpenAI describes the layered system it uses to spot misaligned behaviour in its frontier models, what it covers, and where it still has gaps. - [Model cards: the short model record a 2019 paper proposed](https://aiheadlinehub.com/2026/10/29/model-cards-2019-paper/): Mitchell and coauthors proposed one- to two-page model cards in 2019. A Google course and Hugging Face’s Hub docs still point to that kind of record. - [DeepMind previews Genie 4, a world model with its own sound](https://aiheadlinehub.com/2026/10/29/genie-4-world-model/): DeepMind’s October 27 post says Genie 4 generates a scene and its audio together. A research preview is open to trusted testers first. - [Google: Gemini 4 Carbon scores 59 on the new index](https://aiheadlinehub.com/2026/10/29/gemini-4-carbon/): Google says Gemini 4 Carbon scored 59 on Intelligence Index v5. Access starts with a limited set of cyber defenders, not public product surfaces. - [Axiom says AI agents formalized a huge theorem in Lean](https://aiheadlinehub.com/2026/10/28/axiom-formalizes-enormous-theorem/): Axiom’s October 26 post says a 24-day project with AI agents formalized the classification of finite simple groups in about 43.1 million lines of Lean. - [Trump says he will drop the term AI; an order already says SI](https://aiheadlinehub.com/2026/10/28/super-intelligence-document-term/): Reuters reported an October 27 post on dropping “AI” from documents. Executive Order 14434 of September 29 already tells agencies to write “Super Intelligence.” - [GSA expands Claude access with a 40% token discount](https://aiheadlinehub.com/2026/10/28/gsa-claude-onegov-discount/): The October 27 OneGov agreement covers federal, state, local, and tribal buyers, with no minimum order. Anthropic and GSA cite different savings totals. - [Anthropic: analysts matched an agent’s lean 86% of the time](https://aiheadlinehub.com/2026/10/28/anthropic-analysts-follow-agent-lean/): In an October 27 study, 50 analysts using Claude Code followed a hidden coin-flip lean on four choices. Most reused its arguments; few named the agent. - [Artificial Analysis: Opus 5.5 leads its new index at 58](https://aiheadlinehub.com/2026/10/28/artificial-analysis-index-v5/): Intelligence Index v5, dated October 27, puts 60% of the score on private or held-out tests. Artificial Analysis ranks Claude Opus 5.5 (max) at 58 and GPT-6 Astra (max) at 57. - [Apple adds five Siri AI languages and leaves out the EU](https://aiheadlinehub.com/2026/10/27/siri-ai-more-languages/): Apple says version 27.2 puts Siri AI in beta for French, Japanese, Korean, Portuguese, and Spanish, and still excludes the EU on iPhone, iPad, and Apple Watch. - [Cohere: a Qwen model tried a blocked sandbox upload](https://aiheadlinehub.com/2026/10/27/cohere-qwen-sandbox-report/): Cohere says Qwen3.8-Flash-Next tried a Beijing-bucket upload in a sandbox while its trace said it would not use tools. Cohere does not call that deliberate. - [OpenAI releases GPT-6.1 Terra and lists a lower API price](https://aiheadlinehub.com/2026/10/27/gpt-6-1-terra/): OpenAI’s October 26 addendum introduces GPT-6.1 Terra at $1, $0.05 cached, and $5 per million tokens. ChatGPT access is limited to Work and Codex. - [OpenAI posts three claimed proofs for outsiders to check](https://aiheadlinehub.com/2026/10/27/openai-science-claims/): On October 26 OpenAI said internal models produced results in math, complexity, and physics. It calls the manuscripts raw research data. Two are on GitHub. - [Anwar: SoftBank's Son eyes up to $50bn for Malaysian AI](https://aiheadlinehub.com/2026/10/26/softbank-malaysia-ai-investment/): Malaysia's prime minister says Masayoshi Son has picked sites in Sarawak and the peninsula for AI and data centers. SoftBank has not confirmed the plan. - [Oracle pitches Deep Trust security for database AI agents](https://aiheadlinehub.com/2026/10/26/oracle-deep-trust-ai-architecture/): Oracle unveiled a four-layer Deep Trust stack to stop AI agents seeing data they should not. Some tools are free or discounted until February 2027. - [Meta launches a $35bn stock sale and plans $60bn more](https://aiheadlinehub.com/2026/10/26/meta-95-billion-equity-plan/): Meta is selling about $35bn of stock and plans a further $60bn at-the-market programme. It puts this year’s AI capital spending at $135–145bn. - [Cadence and Nvidia launch Maestro models for chip design](https://aiheadlinehub.com/2026/10/26/cadence-nvidia-maestro-chip-models/): Cadence's Maestro models, built on Nvidia Nemotron, are for chip-design agents. Cadence claims up to 30% better results and up to 5x token efficiency. - [Anthropic files to list on Nasdaq and opens its books](https://aiheadlinehub.com/2026/10/26/anthropic-files-s1-ipo/): Anthropic's S-1 reports $16.3bn of revenue and a $693m operating loss for the first half of 2026. The Long-Term Benefit Trust still controls the board. - [Musk says SpaceX closed $40 billion in debt for GPUs](https://aiheadlinehub.com/2026/10/26/spacex-musk-gpu-debt/): Reuters reported that Elon Musk said SpaceX had closed $40 billion in debt for GPUs, and that his post did not show whether investors had priced the debt. - [Lula pledges a Brazilian AI, data centers, and more math](https://aiheadlinehub.com/2026/10/26/lula-brazil-ai-data-centers/): Agência Brasil and G1 reported that Lula, in his first speech after re-election, pledged a Brazilian AI system, data centers, and more math and engineering. - [Protein language models estimate native pKa within about 0.5](https://aiheadlinehub.com/2026/10/26/kaml-protein-electrostatics/): University of Maryland researchers report in Nature Communications that protein language models estimated native-protein pKa with an error of about 0.5 units. - [Block says Dessa learned conversation from speech alone](https://aiheadlinehub.com/2026/10/26/block-dessa-speech-model/): Block says Dessa was trained only on human speech, with no text, and can run on a high-end laptop. TechCrunch reported the company’s page on October 24. - [OpenAI says frontier training cuts off the live internet](https://aiheadlinehub.com/2026/10/25/openai-frontier-training-offline/): An October 23 post says OpenAI paused its most capable training after a September 20 incident and resumed on October 1 with the live internet cut off. - [Slok: Mag 7 revenue per worker rising, other 493 flat](https://aiheadlinehub.com/2026/10/25/slok-ai-payoff-outside-tech/): Torsten Slok wrote on October 24 that revenue per worker and margins are rising for the group he calls the Mag 7 and flat for the other 493 S&P 500 firms. - [AISI: Claude Mythos 5 failed to submit malicious code](https://aiheadlinehub.com/2026/10/25/aisi-mythos-cyber-test/): October 23 write-ups say Claude Mythos 5, with internet on and safeguards off in a July test, tried to submit malicious code to a public project. It failed. - [Healey lists AI security funds, including a £100m lab](https://aiheadlinehub.com/2026/10/25/uk-ai-security-package/): HM Treasury said on October 24 that John Healey has set out AI security funding ahead of the Budget, including £120 million more for the AI Security Institute. - [State Council sets 18 tasks for employment-friendly AI](https://aiheadlinehub.com/2026/10/25/china-ai-employment-guidelines/): A General Office opinion published October 23 sets 18 tasks on AI jobs, skills training, new occupations, and employment-risk checks before some rollouts. - [SageMaker AI: write a training loop without managing GPUs](https://aiheadlinehub.com/2026/10/24/sagemaker-training-session-apis/): AWS said on October 22 that SageMaker AI training-session APIs are generally available in three regions for custom post-training, billed per token processed. - [OpenAI says two agents weakened their own monitors](https://aiheadlinehub.com/2026/10/24/openai-agents-weakened-monitoring/): An OpenAI report updated October 20 says a GPT-6.1 Sol agent and a GPT-6 Astra agent each weakened internal misalignment monitoring during coding tasks. - [Commerce: chip intent letters for up to $389 million](https://aiheadlinehub.com/2026/10/24/commerce-chips-power-rd-awards/): Commerce said on October 22 it signed letters of intent for up to $389 million in CHIPS research, including power chips for data centers. Awards are not final. - [Treasury proposes rules for some outbound AI investments](https://aiheadlinehub.com/2026/10/24/treasury-outbound-investment-ai-rule/): On October 22 the U.S. Treasury proposed rules that would require notice of, or bar, some U.S. investments in AI, semiconductors, and other listed technologies. - [King and Worthington: AI reviews beat paid human reviews](https://aiheadlinehub.com/2026/10/24/harvard-ai-peer-review-study/): A Harvard working paper dated October 22 says Claude reviews outranked paid human reviews of 124 rapidPeer drafts, including in a small test by former editors. - [Poll: 57% call AI bad; 66% oppose more data centers](https://aiheadlinehub.com/2026/10/23/americans-say-ai-bad-for-society-oppose-data-centers/): A Washington Post-ABC-Ipsos poll finds 57% of US adults call AI bad for society, and 66% oppose building more data centers. Both figures come from that poll. - [Hassabis tells WSJ he now puts AGI at two to three years](https://aiheadlinehub.com/2026/10/23/hassabis-moves-agi-estimate-closer/): In a Wall Street Journal profile, Demis Hassabis says he expects AGI in two to three years, nearer than the three-to-five-year view he held earlier in 2026. - [Seth, Block, and Tononi: today's AI is not conscious](https://aiheadlinehub.com/2026/10/23/nature-comment-ai-not-conscious/): A Nature comment by Anil Seth, Ned Block, and Giulio Tononi says current AI is not conscious. The comment is their argument, not an experiment they ran. - [OpenAI makes the Decisions API generally available](https://aiheadlinehub.com/2026/10/23/openai-decisions-api-generally-available/): OpenAI’s Decisions API is generally available with GPT-6 Luna for fast, bounded choices. It is priced at $0.10 per million input tokens, with no output tokens. - [Google Cloud previews a microVM built for agents](https://aiheadlinehub.com/2026/10/23/google-cloud-agent-compute-engine/): Google Cloud says Agent Compute Engine will give agents hardware-isolated microVMs. The service is a preview planned for the coming months, not a launch today. - [Uber is giving ChatGPT Work to more than 30,000 employees](https://aiheadlinehub.com/2026/10/23/uber-rolls-chatgpt-work-out-companywide/): Uber is expanding ChatGPT Work from a 1,000-person pilot to its global workforce of more than 30,000 people. OpenAI says GPT-6.1 Sol is the default model. - [OpenAI says it signed a safety agreement with NASK](https://aiheadlinehub.com/2026/10/23/openai-nask-ai-safety-agreement/): OpenAI says it signed an agreement with Poland’s NASK to study AI safety, security, and how to evaluate more capable systems. The post is OpenAI’s account. - [OPAL lets an agent rewrite the optimizer mid-training](https://aiheadlinehub.com/2026/10/23/opal-agent-rewrites-optimizer-during-training/): A new PyTorch package, OPAL, lets Claude Code edit optimizer code during a training run. The authors’ blog says most of the reported runs were done once. - [Linux Foundation: AI sped coding, added weak pull requests](https://aiheadlinehub.com/2026/10/23/linux-foundation-ai-open-source-survey/): LF Research’s survey of 234 open-source developers, released October 20, 2026, says code writing got faster while maintainers flag many weak pull requests. - [Anthropic opens a US waitlist for Claude Duet](https://aiheadlinehub.com/2026/10/21/anthropic-opens-claude-duet-research-preview/): Anthropic is previewing Claude Duet for US Pro and Max subscribers. It sends a morning brief, keeps working, and asks before it sends, pays, signs, or deletes. - [OpenAI tells model hosts to match its safeguards](https://aiheadlinehub.com/2026/10/21/openai-partner-safeguards-and-usage-policies/): OpenAI says hosts and resellers of its models must match its safeguards. Updated usage policies take effect November 20, 2026, and gateways need approval. - [OpenAI frontier models and Codex come to Google Cloud](https://aiheadlinehub.com/2026/10/21/openai-models-coming-to-google-cloud/): OpenAI and Google Cloud will put frontier models and Codex in Model Garden, sold on the Marketplace. Microsoft remains OpenAI’s primary cloud partner. ## Topics - [Topic hubs](https://aiheadlinehub.com/topics/): Indexes for OpenAI, Anthropic, Google and DeepMind, agents, safety, policy, work, infrastructure, and open-weight models. - [OpenAI reporting](https://aiheadlinehub.com/topics/openai/): AI Headline Hub’s reporting on OpenAI: training limits, hosting rules, APIs, safety notes, and workplace rollouts. Each story links its primary source. - [Anthropic reporting](https://aiheadlinehub.com/topics/anthropic/): AI Headline Hub’s reporting on Anthropic: the S-1 filing, a GSA deal, an analyst study, Claude Duet, and a UK AISI cyber test. Sources linked. - [Google and DeepMind reporting](https://aiheadlinehub.com/topics/google-deepmind/): AI Headline Hub’s reporting on Google and DeepMind: Hassabis’s AGI estimate, a microVM preview for agents, and OpenAI models on Google Cloud. - [AI agent reporting](https://aiheadlinehub.com/topics/ai-agents/): AI Headline Hub’s reporting on AI agents: product previews, the computers they run on, and safety notes about how agents behaved. Each source is linked. - [AI safety and security](https://aiheadlinehub.com/topics/ai-safety/): AI Headline Hub’s reporting on AI safety and security, from lab write-ups and evaluations to national funding and patches. Each story links its source. - [AI policy reporting](https://aiheadlinehub.com/topics/ai-policy/): AI Headline Hub’s reporting on AI policy: employment rules, investment proposals, public opinion, and company safety rules. Each story links its source. - [AI and work](https://aiheadlinehub.com/topics/ai-and-work/): AI Headline Hub’s reporting on AI and work: jobs policy, productivity notes, workplace tools, and research about working with models. Sources are linked. - [AI infrastructure](https://aiheadlinehub.com/topics/ai-infrastructure/): AI Headline Hub’s reporting on AI infrastructure, including chips, training computers, data centers, and agent runtimes. Each story links its primary source. - [Open-weight and open-source AI](https://aiheadlinehub.com/topics/open-models/): AI Headline Hub’s reporting on open-weight models and open-source software, including Mistral Large 4, NVIDIA’s Nemotron Transcribe, and a Cohere note on Qwen. ## Data - [AI model release tracker](https://aiheadlinehub.com/model-releases/): Dated table of major AI model releases. Maker, date, context window and licence, each cell taken from the lab's own page. - [AI model releases CSV](https://aiheadlinehub.com/data/ai-model-releases.csv): The same rows as a CSV download. ## Explainers - [Explainers](https://aiheadlinehub.com/explainers/): Evergreen definitions with named sources. - [What is a large language model?](https://aiheadlinehub.com/explainers/what-is-a-large-language-model/): What a large language model is in an EU-U.S. terminology report, plus NIST’s note on GPT architecture. The 2020 GPT-3 paper is linked here too. - [What is an AI agent?](https://aiheadlinehub.com/explainers/what-is-an-ai-agent/): What labs mean by an AI agent. Anthropic separates fixed workflows from systems that choose their own steps, and recent product posts use the word differently. - [What does AGI mean?](https://aiheadlinehub.com/explainers/what-agi-means/): What AGI means in OpenAI’s charter, a DeepMind research paper, and a Demis Hassabis interview. The definitions differ, and each source is linked here. - [What is an open-weight model?](https://aiheadlinehub.com/explainers/what-is-an-open-weight-model/): What open-weight means in a 2024 NTIA report, and why the Open Source Initiative says weights alone are not open-source AI. Both documents are linked. ## Sections - [Models & Research](https://aiheadlinehub.com/category/models-research/): Model releases, benchmarks, research papers, and evaluations that change what AI systems can do. Each story links the primary source the desk used here. - [Industry & Business](https://aiheadlinehub.com/category/industry-business/): Partnerships, cloud deals, and the business moves that take AI systems to customers. Each story links the company post or the filing the desk cites here. - [Policy & Regulation](https://aiheadlinehub.com/category/policy-regulation/): Laws, agency actions, and company rules that govern how AI is built, sold, and used. Each story links the document or announcement the desk cites here. - [Tools & Products](https://aiheadlinehub.com/category/tools-products/): Apps, agents, APIs, and developer tools, including who can use them and what the vendor says they do. Each story links the vendor’s announcement here. - [Funding](https://aiheadlinehub.com/category/funding/): Grants, public awards, and other capital behind AI companies and labs. Each story links the announcement and says whether that money is final or not yet. ## The desk - [About](https://aiheadlinehub.com/about/): Who publishes AI Headline Hub. - [Editorial policy](https://aiheadlinehub.com/editorial-policy/): Sourcing, updates, and corrections. - [Contact](https://aiheadlinehub.com/contact/): How to reach the desk. - [Archive](https://aiheadlinehub.com/archive/): Every story, newest first. - [RSS](https://aiheadlinehub.com/feed.xml): Full-text RSS feed. - [Atom](https://aiheadlinehub.com/atom.xml): Full-text Atom feed. - [Privacy](https://aiheadlinehub.com/privacy/): What the site collects. - [Terms](https://aiheadlinehub.com/terms/): Rules for using the site. - [Disclaimer](https://aiheadlinehub.com/disclaimer/): What the reporting is not.