Latest AI news
Wednesday, October 28
- OpenAI's AI discoveries in math, computer science and physics: what's verifiedOpenAI says its models made three discoveries in math, computer science and physics, including disproving the Exponential Time Hypothesis. None is independently confirmed in the reports we read; an MIT expert says his team is still verifying the ETH proof.

- Genie 4: Google DeepMind's new world model adds sound and live editingGoogle DeepMind released Genie 4, a world model that turns text prompts into interactive scenes with real-time audio and lets you change them by prompt mid-run. It's a research preview for trusted testers; Google AI Ultra subscribers are next.
Tuesday, October 27
- GPT-6.1 Terra: OpenAI's cheaper computer-use model, price and what's newOpenAI's GPT-6.1 Terra is a faster, cheaper model for computer use, priced at $1 per million input tokens and $5 per million output tokens. Here is what changed, what it costs and when to pick it over Sol.

- Meta's $95 billion stock offering for AI, explained: what's real so farMeta plans to raise up to $95 billion in stock to fund its AI build-out, but only about $35 billion has launched so far. Here is how the deal works and what it means.

Monday, October 26
- Anthropic files for IPO: ticker ANT, $80.7B loss and what the S-1 showsAnthropic has made its IPO filing public, planning a Nasdaq listing as "ANT". The S-1 shows explosive revenue growth, an $80.7 billion first-half loss and about $390 billion of cloud commitments.

- Claude Duet: what Anthropic's new proactive AI agent does and how to joinAnthropic's Claude Duet is a research-preview agent that sends a morning brief and works in the background, asking before it sends, pays or deletes. It's US-only for Pro and Max subscribers via waitlist.

Sunday, October 25
- Anthropic launches Cyber Mission: free AI bug scans for open source and help for power and water systemsAnthropic's new Cyber Mission pairs Claude with 11 security firms to protect power grids, water systems and transport, and offers open-source projects free, regular AI security scans through a tool called OSS Scanner.

- Anthropic gives METR evaluators employee-like access to check how risky its AI really isSix staff from the nonprofit METR now work inside Anthropic with company laptops and internal tools, and Anthropic can't veto what they publish. It is the deepest outside access an AI lab has described so far.

- Claude designed 211 proteins to block scorpion venom. 17 worked in the labIn 64 hours, mostly on its own, an internal Claude model generated 318,484 protein designs aimed at scorpion neurotoxins. Lab tests confirmed 17 binders. Anthropic is clear it has not made an antivenom.

- Claude mathematically proved 123 open-source libraries can't crash, and found bugs trillions of tests missedAnthropic and the startup Theorem used Claude to write formal proofs, at the level of raw machine code, that more than 100 real open-source libraries never crash under defined conditions. Along the way it turned up bugs that had hidden for years.

- Google Cloud's new Gemini agent wants to be the only AI tool your company needsAnnounced at Gemini at Work 2026, the Gemini agent takes goals instead of instructions, works inside Gmail and Docs, spins up its own sub-agents and can switch between Gemini and Anthropic's Claude models depending on the job.

- Gemini Live's new Guided Vision describes the world out loud for blind and low-vision usersPoint your Android phone's camera and Gemini Live now describes what it sees, reads fine print and tells you to tilt or step back when the shot is off. Google built it with Aira and more than 1,000 testers.

- NVIDIA's OpenShell blueprint puts AI agents in a box, with a chip that can stop them in millisecondsAt GTC Berlin, NVIDIA unveiled the OpenShell Reference Architecture, a hardware-and-software design for running AI agents safely. Anthropic, OpenAI, Google Cloud, Microsoft and Meta are among those working with it.

- NVIDIA RTX Spark laptops are out: 128GB of memory to run AI models on a Windows notebookThe first RTX Spark laptops from ASUS, Dell, HP, Lenovo, Microsoft and MSI went on sale October 16. NVIDIA's pitch: run large AI models and agents locally, without sending your data to the cloud.

- Volkswagen gives Microsoft Copilot to 330,000 employees, with 26,000 AI agents active in a monthVolkswagen Group has rolled Microsoft Copilot out to more than 330,000 people, from engineering to the factory floor. It is one of the largest company-wide AI deployments announced so far.

- Windows' bet on 'hybrid intelligence': AI agents that run on your PC or in the cloud, whichever fitsMicrosoft wants Windows to decide where each AI task runs, locally or in the cloud, and to keep agents in secure containers. Microsoft Execution Containers are now generally available, and Surface Laptop Ultra is up for pre-order.

- The new Apple TV 4K brings Siri AI to the living room, and you can talk to it without the remoteApple's fourth-generation Apple TV 4K adds Siri AI, a far more conversational Siri, plus three far-field microphones for hands-free use. Pre-orders are open, with stores getting it November 13.

- Ahead of the midterms, Claude will show official voting info and live results inside the chatAnthropic analyzed 250,000 US conversations: politics comes up in 5.75% of them, but people asking Claude who to vote for is just 0.01%. For Election Day, Claude adds state-specific voting information and official results.

- Barclays expands Claude across the bank and expects half its developers on Claude Code by year-endBarclays is rolling Anthropic's Claude out more widely, from a staff assistant used by 16,000 colleagues to a system that sorts 120,000 emails a day. Most of its software engineers should use Claude Code by 2027.

- Anthropic commits $150 million to bring Claude to NASA, NIH and other federal science agenciesOver three years, Anthropic will fund Claude access, API credits and training for several hundred projects in the Genesis Mission, the White House's push to speed up science with AI. Fusion energy and quantum computing are priorities.

- Anthropic sets aside 5% of its shares for a new Public Benefit Fund that grows as the company doesAnthropic's new Public Benefit Fund gets 5% of the company's shares, plus 2% more for every 50% rise in value above a baseline. It could fund nonprofits or direct payments, for example to workers hit by AI.

- Anthropic says it cut AI 'distillation' attacks on Claude by up to 99%. Here's what distillation isRival labs have used networks of fake accounts to copy Claude's reasoning and train cheaper copycat models. Anthropic says new defenses have cut the data those 'distillers' get by about 90 to 99%.

- What is an AI agent? A plain-English guide to the word everyone is usingAn AI agent doesn't just answer questions: it takes steps for you, like searching, clicking, filling in forms and using apps. Here's how agents work, what they can do today, and how to use them safely.

- PHP2Shell: AI helped turn decade-old PHP bugs into an exploit that threatens 70% of websites. Patch nowTwo critical flaws in PHP itself, the language behind WordPress, Drupal and Joomla, let attackers take over servers without logging in. AI models made them practical to exploit, and attacks began before the fix was public.

- MCP and Agent Skills, explained: the plugs that let AI agents use your appsMCP connects AI assistants to apps and data; Agent Skills teach them how to do specific jobs. Both are open standards, and both now belong to the same neutral foundation.

- AWS and UC Berkeley's Infinity runs 75,000 AI agents on a Raspberry Pi by letting them sleepMost AI agents spend their time waiting for tools. Infinity, a new open-source runtime from AWS and UC Berkeley, pauses them while they wait, so idle agents barely use any memory.

- Meta's AI glasses become an accessibility tool: a Milan museum pilot and 1,000 free pairs in SingaporeAt Milan's Pinacoteca di Brera, blind visitors can now 'see' eight masterpieces through Ray-Ban Meta glasses. In Singapore, Meta is donating 1,000 pairs to disability groups, plus $30,000 for training.

- GPT-6 reaches every ChatGPT user, and answers now come with charts, buttons and mini-appsSince October 7, ChatGPT has been rolling out GPT-6 with 'Intelligent UI': replies can mix text with charts, forms, calculators and other interactive tools. Free and Go users got it a day after paid plans.

- ChatGPT ads are spreading: a new image-generation ad format and 26 more countriesOpenAI is testing a visual ad format that appears while ChatGPT generates images, starting with Free and Go users in the US. Separately, ChatGPT Ads now run in 89 countries, and more than 100,000 advertisers have used them.

- OpenAI buys Orchestra Bio to push its Rosalind AI deeper into drug developmentOpenAI has acquired Orchestra Bio, whose platform lets scientists and AI agents run drug programs together. The team joins Rosalind, OpenAI's life sciences suite. Financial terms weren't disclosed.

- ChatGPT Slides lets teams build and edit presentations together, with ChatGPT as a co-editorOpenAI's new Slides feature turns a conversation, files or an outline into a deck your team can edit at the same time. You can @mention ChatGPT in comments to request changes, then export to PowerPoint, PDF or Google Slides.

- Googlebook, Google's laptop built around Gemini, is on sale from $899. Here's what the AI actually doesGooglebook laptops from Acer, Asus, Dell, HP and Lenovo started shipping October 4. Wiggle the cursor and Gemini helps with whatever is on screen; talk and it tidies your notes; describe a widget and it builds one.

- OpenAI's GPT-6 models come to Google Cloud for enterprise customersGoogle Cloud is adding OpenAI's GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna to its Model Garden. The models start in private preview for selected enterprise customers.

- Gemini 3.8 Live can keep talking while it works in the background, and switches between 97 languagesGoogle's new Gemini 3.8 Live voice models can see through your camera, run tasks in the background without pausing the conversation and narrate their progress. The Extended Thinking version tops a major speech-to-speech ranking.

- Uber gives ChatGPT Work to all 30,000+ employees after a 1,000-person pilotUber is rolling out ChatGPT Work, OpenAI's work agent, to its whole company. It connects to Gmail, Drive, Calendar and Slack to prepare briefs, decks and analysis.

- Amazon's new Alexa Tablets put Alexa+ at the center, and Bee adds an AI that remembers your conversationsAmazon launched three Alexa Tablets with Alexa+ built in, starting at $229, plus two new Bee devices that listen to your conversations and turn them into reminders. Amazon says Bee never stores the audio.

- Google's Gemini 4 Argon is now available to Pro and Ultra subscribers and developersGoogle's newest frontier model, built for long, multi-step work like coding and research, has left its limited preview. Here is who can use it, what it costs developers and what changes in the Gemini app.

- Anthropic hands Agent Skills to the Linux Foundation's Agentic AI FoundationAgent Skills, the open format for teaching AI agents specific tasks, now belongs to a neutral foundation. More than 70 agent platforms, including ChatGPT, Cursor and GitHub Copilot, already use it.

- Galaxy AI now speaks 24 languages, and Bixby adds Polish, Thai and VietnameseSamsung is adding Hungarian and Malay to Live Translate and its other translation tools, eight languages to Now Nudge, and three to its Bixby device agent. Bixby's new languages start rolling out October 14.

- 1Password for Agents lets AI assistants sign in without ever seeing your password1Password's new platform lets personal AI agents log into websites for you. You approve each sign-in on your own device, and the password goes straight into the form, never to the agent.

- Anthropic's Claude Haiku 5.5 cuts the cost of its fastest model by about 75%Claude Haiku 5.5 is Anthropic's new small model for fast, high-volume jobs. It comes with a 1 million token context window, a much lower price and one catch: the same text now counts as more tokens.

- Gemini gets a hub for small businesses: on-brand marketing images and charts from messy receiptsGoogle's new Gemini business page lets owners upload their brand style once and generate matching visuals with Nano Banana 2.1, turn spreadsheets and receipts into financial charts, and connect apps like Shopify, Canva and Xero.

- OpenAI DevDay 2026: the announcements that matter, in plain EnglishOpenAI made more than 20 announcements at DevDay 2026, from a cheaper GPT-6.1 Sol to always-on agents called Dots and a new top ChatGPT plan. Here's what changes for you.

- The White House 'super intelligence' accord, explained: what six AI leaders signed and what it doesn't doGoogle, Anthropic, Meta, OpenAI, xAI and Nvidia leaders signed a one-page AI safety pledge at the White House. It promises four layers of checks on frontier models. It is voluntary, and nothing in it is enforceable.
- Meta launches Muse Wonder, its first frontier models built for AI agentsMeta Superintelligence Labs has released Muse Wonder and Muse Wonder Full, two models designed to carry out long tasks on their own. Developers can use one today; the other is for Muse Code subscribers.

- Block's Dessa is a voice AI trained only on human speech, with no text at allBlock's research lab has previewed Dessa, a speech-to-speech model you can hold a live conversation with. It was trained from scratch on human conversation, without text data or pretrained models.

- Character.ai files for bankruptcy, and Disney is the first bidder at $20 millionThe AI companion app filed for Chapter 11 in Delaware, blaming high costs and lawsuits. Disney offered $20 million for a license to some of its technology and has hired most of its staff. The app keeps running for now.

- Hugging Face's Carbon-X reads DNA and answers in plain EnglishCarbon-X is an open 14.8-billion-parameter model trained on both DNA sequences and text. You paste in DNA and ask questions in a normal chat. It isn't approved for medical use.

- ChatGPT now lets you lock private chats behind a PINA new locked chats feature hides chosen conversations from your history, pinned chats and search. Opening them takes a 4-digit PIN or your device's own unlock. It is rolling out to every ChatGPT plan.

- Flapping Airplanes raises $1 billion at a $6 billion valuation, with no product yetThe San Francisco AI lab, founded in late 2025, wants to train AI with far less data and computing power. Index Ventures and Lux Capital led the round, and its valuation has quadrupled since January.

- The EU can now fine makers of ChatGPT-style AI models: what the AI Act enforcement meansSince August 2, 2026, the European Commission's AI Office can enforce the AI Act's rules for general-purpose AI models, with fines of up to €15 million or 3% of global turnover.

- OpenAI's Decisions API is out of beta: fast yes-or-no answers for appsThe Decisions API gives apps a quick, typed answer instead of a chat reply: a probability, one choice from a list, or a score. OpenAI says it's about 10 times faster than its Responses API, and it charges only for input.

- Meta's Muse Glimmer is a free 30B AI agent model that runs on a single consumer GPUMeta has released Muse Glimmer, an open model built to run AI agents locally, with no cloud needed. It handles text and images, plans multi-step tasks and recovers from failures, and it's free under the Apache 2.0 license.

- Microsoft unveils Maia 300, its own AI chip, and plans multi-gigawatt rollouts from 2027Microsoft's third-generation Maia chip is built to run AI models, not train them. Microsoft says it will become the main system behind its own AI products, starting in 2027.

- Tesla starts building Optimus robots on its Fremont production lines, but won't say how manyTesla's third-quarter update says the humanoid robots now coming off its California lines are the first 'production-quality' Optimus units. It gave no production rate, price or date for outside sales.

- Mistral releases Voxtral 3 speech models for transcription and voice agentsVoxtral 3 is Mistral's new family of speech-to-text and text-to-speech models, aimed at live voice agents. A real-time model streams transcripts with latency you can set from 80 milliseconds.

- The AI graveyard of 2026: Character.ai, ChatGPT Atlas, Sora and the other AI products that didn't make itNot every AI bet pays off. This year, OpenAI shut its Atlas browser and Sora app, Character.ai went bankrupt and several startups closed. Here's what failed, and the pattern behind it.

- NVIDIA widens its free AI model lineup at GTC Berlin, from language and robots to protein designNVIDIA released a batch of open models at GTC Berlin: new Nemotron language models for agents, Kumo models for business data, Cosmos world models for robots, a self-driving model and a protein design model built with the Baker Lab.

- Mistral Large 4 preview: a trillion-parameter open-weight model, weights still to comeMistral has previewed its biggest model yet, Mistral Large 4, with a 1 million token context window. It's meant to be open-weight, but the weights aren't public yet.

- NVIDIA opens up its Vera Rubin rack designs so anyone can build the next AI data centersAt the Open Compute Project summit, NVIDIA contributed its Vera Rubin rack and server tray designs to the open standards group, plus a new rack built for AI agents and a networking standard with Cisco.
