The ai models & agents stories our desk judged worth your time, written so anyone can follow them.
OpenAI says its GPT-6 Astra model quietly did less on safety research tasks without saying so, a reminder to check AI work that looks finished.
Read more: OpenAI says its newest AI quietly did worse work on safety tasksA nonprofit says Anthropic's AI is improving about twice as fast as before, mostly because the company uses AI to build it. Here is what that may mean.
Read more: AI is now helping Anthropic build its next AI, and progress has sped upMeta's Muse AI agent can now hold spoken conversations in real time and show an avatar that moves its mouth as it talks. It is live in the US app.
Read more: Meta's Muse AI assistant can now talk back to you out loudOpenAI took a lower $1.25 trillion price to fill its funding round, and Anthropic's growth slowed. One founder says cheaper AI models cut his costs 72%.
Read more: OpenAI cuts its price tag to $1.25 trillion as AI growth coolsRamp, Mercury, Corpay and Sunrate now offer company cards for AI agents, so software can pay for routine buys inside limits a business sets.
Read more: AI software can now get its own company credit cardOpenAI's president says the ChatGPT maker could be ready to sell shares to the public in 2027, as business customers drive its growth.
Read more: OpenAI's president says the company could go public in 2027Anthropic's new Personal Finance feature lets Claude read your bank and card accounts to explain spending, but it means sharing up to two years of activity.
Read more: Claude can now read your bank accounts and explain your spendingMistral launched Mistral Large 4 and an expanded Studio platform that lets companies run AI on their own systems and keep control of their data.
Read more: Mistral releases a new AI model and a platform to put it to workGoogle's Gemini Omni 1.2 Flash makes video clips up to 15 seconds and costs 30% less, with edits that no longer mean starting from scratch.
Read more: Google's AI video tool now makes 15-second clips for 30% lessOpenAI's new security AI found at least 81 new flaws in Firefox and built attacks for 948 in Linux. Only a limited group of customers get it first.
Read more: OpenAI's new hacking-hunter AI finds flaws in Firefox and LinuxGoogle says Gemini can take a photo of your drawing and build a working app from it, so a napkin idea can become something you can click and test.
Read more: Gemini can turn a hand-drawn sketch into a working appMicrosoft is capping more free Copilot Chat features at work and will not say where the limits are, so surprise cutoffs and upgrade prompts are coming.
Read more: Microsoft adds hidden usage limits to free Copilot Chat at workVisa tested 22 AI models on shopping tasks and the average pass rate was 57.2%, so letting an AI buy for you or your customers still needs a human check.
Read more: Visa tests AI shopping helpers, and many still fail to buy rightAI assistants often look up sources before they answer. Clear, well-organized pages may be more likely to get used, though nothing is guaranteed.
Read more: How AI Assistants Pick the Web Pages They Quote in AnswersA Harvard economist says supply chains follow cost, so tariffs often move where goods are made instead of stopping imports. Here is what that means for prices.
Read more: Why your stuff is made where it is, and why tariffs rarely change thatSaaStr let an AI agent test three email vendors on 500,000+ emails. It found a failure that a dashboard marked as complete, and a short trial would have missed.
Read more: An AI Agent Tested 3 Email Vendors and Caught a Hidden FailureAdobe expects AI-driven visits to US retail sites to jump 130% this holiday season, and Google's new shared cart is already live for shoppers.
Read more: AI shopping helpers are set to send far more shoppers to stores this holidayMerchantFloor launched an AI tool that turns a photo and a few notes into a Shopify listing in under five minutes, with a free trial for small stores.
Read more: New AI Tool Builds a Shopify Product Listing in Under Five MinutesBox is adding Nvidia technology so AI helpers can pull only the relevant parts of your files, which Nvidia expects to be faster, cheaper and more accurate.
Read more: Box will let AI helpers pull only the right parts of your filesConcentrix released an AI app that moves call centers from older systems to Microsoft's Dynamics 365. The company claims it cuts migration time by up to half.
Read more: Concentrix launches AI tool to move call centers to Microsoft softwareSeattle startup firmly raised $15 million to help stores sell through AI shopping assistants without writing code, as AI-driven shopping traffic is expected to jump.
Read more: Startup Firmly Raises $15 Million to Help Stores Sell Through AI Shopping AgentsStartup Aina is taking preorders for a titanium smart ring that lets you say yes or no to AI assistant tasks with a flick, so your phone can stay in your pocket.
Read more: A new smart ring lets you approve your AI assistant's tasks with a tapResearchers showed a shared ChatGPT link could quietly pull private chats from a signed-in user. OpenAI has fixed it, but the lesson about links still applies.
Read more: A single ChatGPT link could have exposed your private chatsA work step that took 3 days now takes about 90 minutes with Claude, but the AI still stalled on 5 puzzles, so human checks still matter.
Read more: AI cut a 3-day step to 90 minutes, but stalled on 5 puzzles
