DeepSeek reportedly plans to lead funding for Chinese robotics AI startup Winverse at a $400 million valuation, signaling a broader push into physical AI. The post DeepSeek Eyes $400M Robotics AI Startup in New Funding Push appeared first on TechRepublic.

Apple has launched its most serious push yet into your home, built around its attempt to normalize always-on AI with Siri. As widely expected, the new products include its much-anticipated Apple HomeView smart display, a new Apple TV 4K with Siri AI, the next generation HomePod mini 2, and Apple Home Collection, a range of home accessories designed in collaboration with third-party partners. What is HomeView?  HomeView is a central control and entertainment device for smart homes. Powered by Apple Intelligence with Siri AI, the wall- or stand-mounted smart display costs $299 and has an A18 chip that runs “HomeView” software. Apple confirmed most of the pre-event speculation, most particularly that the product will not be available in the European Union, as Siri AI is currently unavailable there due to restrictions within the Digital Markets Act. Apple did say it is “working constructively with the EU in an attempt to find a path forward.” We’ll see how that turns out. Features include a custom-engineered acoustic system, FaceTime, face recognition, and Apple’s promise on privacy (which is that it provides you with some). The device supports up to six users; it uses face recognition to identify a user and then adapts what they see so they can access their own personal content on the device. That means you’ll see your emails, while your partner will see their own, for example. HomeView is also a new breed of contextually driven AI devices, so it can call on your personal context to answer questions or requests. Wired called it a “giant Apple Watch for the home.” Industry watchers think it looks like a mini iMac G4, with a rounded base and a display on a stalk. same same. pic.twitter.com/etlyfk8bBf — Basic Apple Guy (@BasicAppleGuy) October 13, 2026 It has a 7-inch Liquid Retina display connected to a domed speaker base by a stainless steel hinge that provides a smooth tilt for comfortable viewing on tabletops of any height. It is available in silver, starlight, rose pink, and space gray. It has a built-in speaker reportedly capable of making a great sound, is AirPlay compatible, and works with compatible home security cameras to become an AI-powered home security system.   It’s thought that Apple originally hoped to introduce HomeView a year or so ago but was stymied by its failure to develop Siri AI to schedule. All the same, it’s an important product, as it shows the company experimenting with ways to provide ambient AI experiences that actually benefit consumers. It puts the company up against products in the ilk of Amazon’s Echo Show or Google’s Nest. It also echoes Samsung’s SmartThings ecosystem, but platform integration and privacy may give it the edge. Apple Apple Home Collection partnerships The Apple Home Collection is a rare excursion into partnership, with Apple working with both LG Electronics and Schneider Electric to introduce a new range of smart home accessories developed in collaboration.  It’s unclear whether this is a new licensing system for smart home products, building on the Works with Apple Home accessory branding Apple has offered for a while.  This is smart home tech done better, however, as the big selling point is that you can install these products easily, “starting with one tap with iPhone.” You don’t need to install other apps, create new accounts or otherwise create additional cruft. The products are also designed for privacy, with Home app data encrypted end-to-end, so LG Electronics, Schneider Electric, and Apple can’t read it. They support the Matter and Thread home networking standards. “Apple Home Collection makes it easy to choose smart home accessories with confidence,” said Matthew Costello, Apple’s vice president of hardware engineering. The new gadgets are available in a limited number of nations — mostly the US, UK, and Canada, though some are North America only. As anticipated, the new products include: A video doorbell Floodlight, outdoor, and indoor cams that are iCloud+ compatible Smart deadlock Smart thermostats Apple reached a deal with ADT to offer professional installation of these accessories for a fee, though you don’t need to do this. Prices seem to hover around the $200 mark. The LG Video Doorbell Wired starts at $250 (US), while the Schneider Home Smart Thermostat starts at $200 (US), for example. New Apple TV 4K and HomePod mini The HomePod mini 2 smart speaker is available in five colors and brings support for Siri AI (where available). It uses a new S10 chip to analyze music in real time to optimize dynamic range, loudness, and acoustic performance. A new feature called Sound Glow “dances” to the music, glowing and pulsing color inspired by the album cover. It works as a smart home controller and costs $149. If you’re feeling flush, you can invest in four of these to create a quadrophonic speaker system for use with the new Apple TV 4K. The upgraded Apple TV 4K has a new A19 processor, while built-in far-field microphones bring support for Siri AI where it is available. You can link it to the HomePod mini for better sound. The device works as a home hub and can also access your home video surveillance feeds. The processor upgrade means you’ll get smoother navigation and faster response. The system, which costs $299, also has a new, dedicated AV1 decoder, which should enable more efficient, high-quality streaming while using less bandwidth. An interesting note: Siri AI will be available as a beta on older Apple TV 4K models (second generation or later) in most English-speaking countries.  For the rest of us? It seems relevant to mention a few yet-to-be-determined questions and rumors, such as what support the new Apple home ecosystem will provide for non-Apple devices. For example, if you are the only person in your family using an Android phone, you are locked out of the family experience of this smart home integration. How can that user use the product? When it comes to privacy, does the new Apple TV use Automatic Content Recognition to any extent, and how much viewer information does it collect? One more thing that has been picked up is that Siri AI will be available as a beta on older Apple TV 4K models (second generation or later) in most English-speaking countries.  Now please subscribe to my daily, human-curated Apple-related news headline feed at The Core, or follow me on BlueSky, LinkedIn, or Mastodon.

NVIDIA is making the full NVIDIA Run:ai platform available as open source software, giving customers, developers, and ecosystem partners access to... NVIDIA is making the full NVIDIA Run:ai platform available as open source software, giving customers, developers, and ecosystem partners access to production-grade AI workload and GPU orchestration capabilities. This release expands the NVIDIA DSX AI Factory Platform by establishing NVIDIA Run:ai as an open foundation that enables the broader ecosystem to bring differentiated solutions to market… Source

Meta's Muse Wonder AI models challenge OpenAI and Anthropic with new agent capabilities. Explore benchmark results, pricing, access restrictions, and privacy risks. The post Meta’s New Muse Wonder Models Challenge OpenAI, but Access Comes With Limits appeared first on TechRepublic.

DC Studio / Shutterstock Generative AI can explain a difficult concept, solve a problem, correct a sentence or produce an answer in seconds. That creates enormous opportunities for learning. People have more access than ever before to information about different learning strategies. But the strategies that feel easy or familiar are not necessarily those that produce the most durable learning. You are in charge when you use an AI so it’s a good idea to come up with a learning strategy that works. Understanding this matters because successfully completing a task is not the same as learning from it. AI might help a person produce the right answer today without increasing their ability to independently produce a correct answer tomorrow. Language learning provides a useful example. An app such as Duolingo can provide structure, repetition and motivation. A learner can deepen that experience by trying to produce the answer from memory before looking at or selecting the options provided by the app, repeating unfamiliar sentences aloud or revisiting difficult vocabulary. Deliberately retrieving information can improve longer term retention compared with simply studying material again. A large language model such as ChatGPT can play a different role. It can explain things, compare sentence structures, generate further examples or help investigate something the learner has noticed. When an app is used together with a large language model, they can create a richer learning ecosystem. A small case study involving Duolingo users, for example, found that adding reflective journalling – writing down thoughts and goals – gave learners opportunities to articulate their strategies, difficulties and developing awareness of their own learning. Consider two people using the same AI model. One gets stuck and asks for the answer. The other attempts the problem first and then asks why their answer was wrong. They might request a clue, ask for another example, or explain what they think is happening and ask the AI to challenge them. Both have access to the same tool. But the learning activity is very different. Changing how we use AI Psychologists use concepts such as metacognition – the ability to examine, understand, and regulate your own thought processes – to describe how people monitor and manage their own learning. Effective learners make judgements about what they understand, what needs more practice, which strategies are working and when they need help. Emerging research suggests these capacities also shape how people engage with AI. In one semester-long study of university students using ChatGPT for English-language writing, metacognitive awareness was associated with different patterns of AI use. Interview data showed behaviour ranging from simply copying ChatGPT’s output to using its feedback more actively to improve written work. There’s evidence that AI use can lead to dependence that erodes natural learning skills. Tada Images / Shutterstock The design of the technology matters too. In a large field experiment involving nearly 1,000 secondary school maths students, the researchers gave one group of students access to a standard version of GPT-4. They gave another group of students access to a different version of GPT-4 that was designed to behave like a tutor, rather than simply supply answers to prompts. The students with access to the standard version of GPT-4 saw their performance improve. Yet, when the researchers took their access away, these same students performed 17% worse than students who had no access to AI. The group who studied with the GPT-4 version designed to tutor largely avoided this decline. The tutor group is important because the lesson is not that learners should be denied assistance or simply made to struggle. Effective support can remove unnecessary barriers while preserving the mental activity learning requires. AI can give an answer immediately. But it can also offer a hint, ask someone to attempt the problem first, identify a misconception, generate another example or ask the learner to explain their reasoning. The important question is therefore not simply whether AI is being used. It is which learning activities it is allowing someone to stop doing, and which it is helping them do better. Self-aware learning AI literacy matters. Learners need to understand what these systems can and cannot do, how to question their outputs and when their use is appropriate. But we might also need to think more explicitly about learning literacy. Learning literacy is not a new psychological construct, but a public-facing way of describing the knowledge and judgment people need to regulate their learning in increasingly technology-rich environments. AI literacy asks whether we understand the tool. Learning literacy asks whether we understand learning well enough to decide how that tool should be used. There is good reason to think AI could form part of the solution. A recent meta-analysis of 32 studies involving 3,029 participants reported substantial positive effects of AI-supported interventions on self-regulated and self-directed learning. However, the effects varied according to how the technology and learning experience were designed. That means responsibility cannot sit with learners alone. Schools and universities can make learning itself more explicit, helping students understand why strategies such as retrieval, reflection and feedback work. Developers can design AI that encourages attempts before assistance, offers hints rather than solutions where appropriate, and asks learners to explain or revisit their reasoning. And learners can ask a deceptively simple question: what am I actually learning by using AI this way? We are spending enormous effort making AI better at answering questions. The educational challenge is to become equally good at recognising when those answers are helping us learn, and when they are simply helping us finish. Paul Jones does not work for, consult, own shares in or receive funding from any company or organisation that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.

Benchmarking NumPyro, affine-invariant MCMC, and dynesty against exact ground truth and watching them fail. The post Three Samplers, One Posterior: NUTS vs. Emcee vs. Nested Sampling appeared first on Towards Data Science.

New market research from Counterpoint Research shows that global smartphone sales fell by 20% in the third quarter of this year compared with the same period a year ago. This is the lowest level for a third quarter since 2012. The decline is attributed to the ongoing shortage of memory components, where sharply rising prices for DRAM and NAND memory have driven up manufacturing costs. Manufacturers, in turn, have raised prices for customers. According to Counterpoint, global smartphone prices have risen by about 20% in 2026. In terms of market share, Samsung continues to hold the top spot with 23%. Apple is close behind with 22% of the market, the highest share the company has ever held in a third quarter. Counterpoint does not expect to see any clear recovery in the mobile market for the remainder of the year.

Families are paying hundreds of thousands of dollars a year in long-term care costs. kitiwan mesinsom/Getty Images Business Insider is seeking reader responses to document elder care costs in the US. Caring for elderly relatives can cost families over $200,000 annually. Families are facing uncertainty in predicting long-term care expenses for elderly relatives. Caring for an older relative is expensive. We want to document what long-term care really costs American families, bill by bill. One of the biggest problems families face is not knowing how much they are likely to spend, and there are few resources that offer any frame of reference. We are inviting readers to share their stories with Business Insider reporters to help us assemble a comprehensive look at these costs. We want to know how much caring for an older loved one has cost you and your family. Getting your input will help us break down barriers to accessing financial information and better prepare our readers for their futures. This matters because long-term care costs have burdened families across the country. Some told Business Insider that they've spent over $200,000 a year on memory care facilities. Others have shelled out $450,000 on tiny homes for their parents. Many have put their jobs, retirement planning, and livelihoods on hold to care for their older loved ones, some moving abroad to afford it. To participate in our series, which we anticipate will be published over the course of the next year, please complete the questionnaire below, which asks a few basic questions about your elder care situation. After you do that, a reporter may be in touch with you via email or text about setting up a time to talk through more granular questions and, later, reviewing documents (such as bills, receipts, insurance information, etc.) to verify the costs you report. Any verifying documents will be transmitted via a secure portal. Only Business Insider reporters and editors, as well as employees and service providers that support this effort, will have access to what you share. Business Insider may use your responses, and any information you later provide in follow-up requests, in its reporting on elder care costs. Loading… Be sure to press "submit" on the form above. If you filled out the survey, we thank you very much and look forward to talking with you further. Read the original article on Business Insider

<blockquote cite="https://ghuntley.com/agency/"><p>If the agent isn't using my diagnostics library, I could introspect and update AGENTS.md, but there's still a delta, and it'll drift. Instead, I write a pre-commit hook that checks the git diff for direct calls to observability modules that bypass my diagnostics module. Boom: fail the commit, inject a prompt telling the agent to use the diagnostics library. All of a sudden it's ten out of ten.</p></blockquote> <p class="cite">&mdash; <a href="https://ghuntley.com/agency/">Geoffrey Huntley</a>, nobody should type code by hand anymore. here's what's left.</p> <p>Tags: <a href="https://simonwillison.net/tags/coding-agents">coding-agents</a>, <a href="https://simonwillison.net/tags/generative-ai">generative-ai</a>, <a href="https://simonwillison.net/tags/git">git</a>, <a href="https://simonwillison.net/tags/agentic-engineering">agentic-engineering</a>, <a href="https://simonwillison.net/tags/ai">ai</a>, <a href="https://simonwillison.net/tags/geoffrey-huntley">geoffrey-huntley</a>, <a href="https://simonwillison.net/tags/llms">llms</a></p>

The author is French, but raising her 9-year-old in America. Courtesy of Virginie Romary My 9-year-old daughter is enrolled in an American public elementary school. The school system here differs significantly from the one in France. I like that it is less rigid, with many extracurricular activities and a strong sense of community. I moved from France to the US 10 years ago. My 9-year-old daughter was born in Florida, and when it was time to enroll her in school, I chose not to enroll her in a French school. I wanted her to experience the American education system to be part of the community. I like the school spirit here What I appreciate most about American schools is that they function as communities, whereas in France, they are primarily places for academic instruction. I admire how school spirit is actively promoted here. My daughter takes great pride in her school, and her favorite T-shirt features its colors. She eagerly participates in spirit weeks, dress-up days, and various entertaining events like the color run that she never wants to miss. I also value the morning rituals in the US, including the morning announcements and the Pledge of Allegiance. Such practices do not exist in France. My daughter feels deeply engaged with her school. She has a sense of belonging that I never experienced as a child, because my schooling focused solely on formal instruction, with no fun activities. Americans' school days are shorter At first, when I realized that my daughter's school finishes at 2 pm, I panicked. How was I supposed to work at the same time? However, once I organized our daily routine, I really appreciated that she had a more compact day and more time in the afternoons to focus on extracurricular activities like swimming or playing music. Typically, in France, school days are longer, running from early morning to late afternoon, with more breaks and a longer lunch period. Wednesdays are half days or off days for the youngest, giving children time to participate in extracurricular activities they may want to take part in. However, there are no school-sponsored extracurriculars, and theater, music, or competitive sports are played in clubs outside the school system. Often in the evening, I hear music and cheering from the football matches at the nearby high school, and I look forward to my daughter having this kind of experience. Until then, I value that her days are more balanced, with afternoons dedicated to non-academic skills, even if it means she has fewer holidays throughout the year. School is less rigid in the US In France, I would never see my daughter's teacher in her sleepwear for Pajama Day or her principal wearing bunny ears for Halloween. The boundaries between school and home are stricter there, and teachers tend to adopt a more authoritarian position. In the US, my daughter has a less formal relationship with her teacher, knowing details about her personal life. Her teacher is easily accessible outside class and puts significant effort into encouraging and praising her students. I'm often surprised when my daughter receives prizes, as this practice is not common in France. Her teacher always finds something positive to express, even if it is as simple as saying, "I like your energy," for a child who struggles to remain seated. The French approach is different, and teachers are less inclined to compliment. They focus more on correction and improvement, using a 20-point grading scale that makes a perfect score almost impossible. I found the letter grade system here less stressful, with an A being more attainable. Consequently, making the honor roll is easier. This recognition doesn't even exist in France, where children's achievements are not celebrated publicly. However, not everything is perfect. I also have a few concerns about the American school system, such as the culture of standardized testing and the competitive atmosphere. Still, my daughter enjoys her school, which is a great relief. Read the original article on Business Insider

Mykhailo Fedorov, Ukraine's former defense minister, has developed a new interceptor drone. Kyodo via Reuters Connect Mykhailo Fedorov, Ukraine's digital-focused former defense minister, has unveiled a new interceptor. The interceptor is in testing and has already downed Russia's fastest jet-powered drone, the Geran-5. Fedorov told Business Insider this week that he plans to be producing hundreds a month by the end of the year. Mykhailo Fedorov, Ukraine's former defense minister who pushed for rapid innovation in war technology, has unveiled a new interceptor designed to take down Russia's fastest attack drones. The new interceptor drone is currently flown by a human pilot, but Fedorov said in comments shared with Business Insider that he intends to make it autonomous in flight using artificial intelligence. He aims to be producing hundreds of interceptors a month by the end of the year. Ukraine is increasingly turning to AI-enabled autonomous drones and robots that can continue fighting when electronic warfare severs the connection between systems and operators. The interceptor that Fedorov unveiled is the latest model developed by Ukrainian engineers to combat the rise in Russian jet-powered drones, weapons which are much faster than their propeller-powered predecessors and pose a significant challenge for Kyiv's air defenses. It also reflects a broader shift toward interceptors equipped with artificial intelligence that can fly with less human input. Ukrainian defense firms have developed several new interceptors capable of engaging Russia's fastest drone, the Geran-5, which is powered by a mini turbojet engine and able to reach speeds of up to 375 mph. It's unclear how fast the interceptor can fly, but in recent combat tests over Kyiv, a specialized unit of Ukraine's National Guard used it to take down two Geran-5 drones. Fedorov said that testing is ongoing. The next steps are to improve the technology and scale up production. In translated remarks, he said the interceptor "will be very affordable." "That was our goal from the outset: to develop a low-cost, mass-producible solution that can be scaled rapidly," he said. A view of a Russian Geran-5 captured by Fedorov's interceptor. Mykhailo Fedorov/Screengrab via X During his six months as Ukraine's defense minister earlier this year, Fedorov pushed for a sweeping modernization of the military, including through a robotic-centric approach to combat. His tenure coincided with a string of battlefield successes for Kyiv, primarily through drone warfare. Following his sudden departure in July, Fedorov launched a defense company focused on developing warfighting technology to strengthen Ukraine and support international partners. He said in early September that the first investor will be Alex Karp, the CEO of Palantir, an American AI, data, and operational intelligence firm. On Monday, Fedorov said he intends to invest more in research and development and the Ukrainian engineering team behind the interceptor, making its intellectual property a "key asset" in his portfolio. He also plans to open investment to other partners; the first funding round is expected soon. The development of the new interceptor drone comes as Moscow has ramped up its use of jet-powered drones this year. Ukraine's air force said Russia launched 350 drones in June, but that number swelled to 2,800 in August. These attacks present a serious challenge for Ukraine. Officials said that the interception rate of Russia's propeller-powered drones, which reach 115 mph, typically sat above 90%. Ukrainian President Volodymyr Zelenskyy said in late September that Kyiv shot down just 55% of Russia's jet-powered drones that month. Russia has developed several jet-powered drone models. The Geran-5 more closely resembles a cruise missile than its slower delta-wing predecessors, and it is playing a growing role in Russian strike tactics. Zelenskyy warned last month that Moscow plans to invest heavily in these weapons next year. Fedorov said on Monday that he's "confident that Russia can be stopped" and that Ukraine has already demonstrated it can engage the slower drones. "Our next challenge is to do the same with jet-powered versions," he added. Read the original article on Business Insider

🧠 What decision models are, where product teams are already using them and where they fall short. Real world examples from Vercel, Rippling, Harvey and more.

Apple has unveiled HomeView. Apple Apple's first smart home device is here. The Apple HomeView is a new device with a screen and speaker that acts as a control hub for your connected home devices. It's the latest attempt by Apple to make its ecosystem more "sticky," or harder to leave. Apple has officially entered into a new product category: the smart home. The company on Tuesday announced the HomeView, a new AI-powered hub designed to reside in your home and control your connected home accessories. The device consists of a 7-inch Liquid Retina display that sits atop a domed speaker base and works with Siri AI. The device can be placed on a tabletop, and a separate model can be mounted to the wall. "Apple HomeView is designed to be shared by everyone in the home yet feel completely personal to each user," said Greg Joswiak, Apple's senior vice president of Worldwide Marketing, said in a statement. The product, he said, "recognizes your face to shift from a household screen to your own experience, fills the room with incredible sound, keeps you connected to the people who matter most with FaceTime, and makes it easy to control your smart home." HomeView can recognize up to six people and surface personal content for the user in front of it, such as messages, calendar information, and apps. It draws on "personal context from your iPhone to help you find what you need and get more done, all while keeping everyone's information private and secure," Joswiak said. The device is priced at $299, with preorders opening Tuesday ahead of a November 13 launch. HomeView is priced at $299. Apple Part of Apple's new smart home lineup HomeView anchors Apple Home Collection, a new lineup of smart home cameras, thermostats, and a lock that Apple developed in collaboration with LG Electronics and Schneider Electric. As a smart home dashboard, the device can display camera feeds, adjust thermostats and lights, lock doors, and close blinds using touchscreen or voice commands. Apple said it can livestream up to nine video feeds at once. "With a supported iCloud+ subscription, HomeView works with compatible cameras to deliver intelligent, end-to-end encrypted recognition of people, familiar faces, pets, packages, and vehicles, notifying the household the moment someone arrives at the front door," Apple said. HomeView marks the second major new device announced under new Apple CEO John Ternus, who took over for Tim Cook in September ahead of the company's iPhone event. At that event, Ternus also unveiled the first folding smartphone, the iPhone Duo, which begins preorders later this week. Apple has long supported the smart home ecosystem with its HomeKit standard that third-party device makers could choose to use to help their devices more seamlessly integrate with the iPhone. The HomeView underscores that Apple views the smart home market as one worth competing in with its own hardware. If the new device takes off, a consumer moving outside the Apple ecosystem — or "breaking up with Apple" for Amazon or Samsung or Google or another rival — becomes a harder decision. Apple is famous for deeply integrated software and hardware. This is a developing story… Read the original article on Business Insider

Co-author: “The unique geometry of the valves is the key; they act as collapsible pistons" to deliver venom.

The company’s new releases point to an AI market becoming more personal, with agents and models completing tasks, but enterprises still aren’t ready for this new phase of AI.

Real repositories, seeded agent decisions: what 50,400 runs show about wasted turns, patch churn, and which harness rules help The post Coding Agents Can Solve the Task — But How Much Work Did They Waste? appeared first on Towards Data Science.

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. Open models have real benefits. Do they outweigh the risks? Anthropic recently released a report arguing that a powerful new model released by the Chinese company Z.ai, poses a major cyberattack…

Higgsfield brings creators longer videos with GPT-6 Astra, precise visuals with GPT-Image-2.5, and a path to voice filmmaking with GPT-Live.

A Baltic tech company tucked away on the site of a former sock factory has soared to become Lithuania’s latest unicorn, positioning itself as the workhorse of the data-driven artificial intelligence revolution.

Microsoft has launched Decision-1, an AI model that returns probabilities for predefined answers, allowing software applications to use its outputs directly instead of processing a free-form response from a large language model (LLM). The model targets tasks such as classifying requests, prioritizing incidents, and checking AI-generated responses. Microsoft says it can handle these decisions with lower costs and latency than approaches that rely on general-purpose LLMs. “Instead of writing a free-form response, it takes a fixed set of answer options and returns a calibrated probability for each one in a single pass,” Microsoft said in a blog post. Microsoft-Decision-1 is available through Microsoft Foundry and OpenRouter, according to the company. Microsoft built the model by post-training Qwen3.5-9B and said it plans to rebase it on other models, including Microsoft AI (MAI) and OpenAI models. Model targets specific decisions Decision-1 supports yes-or-no judgments, choices among predefined options, ratings, and rubric-based assessments of AI responses and agent actions, Microsoft said. Developers can use its outputs to classify inputs, rank work, verify responses, or determine the next action in an AI agent workflow. Unlike a general-purpose LLM, which can generate explanations and other natural-language content, decision models return probabilities for the available answers. Microsoft said its Decision-1 model can also evaluate multiple decisions in parallel within a single request. Microsoft cites benchmark results According to Microsoft, Decision-1 achieved the highest accuracy in a comparison of 36 benchmarks covering nearly 150,000 questions kept blind from training. The company also reported that the model was about 2.5 times faster than the runner-up and 35 times faster than GPT-6 Sol. Microsoft lists the price at 4.2 cents per million input tokens, with no charge for output tokens, according to the blog post. In an incident-prioritization example, the company said a request using 86 input tokens and two output tokens cost approximately $0.0000036. At that rate, $1 would cover more than 275,000 similar decisions. The benchmark results and cost calculations are Microsoft’s claims and have not been independently verified. Charlie Dai, vice president and principal analyst at Forrester, said specialized decision models could reduce the cost and latency of agentic AI workflows. “Routing, classification, and approval checks often represent a large share of agent interactions, so a specialized decision model can significantly reduce cost and latency while reserving general-purpose LLMs for reasoning and content generation,” Dai said. Amit Kumar Jena, development manager for AI at Kanerika, said moving routine routing and approval checks away from large generative models represents a meaningful architectural change for enterprises scaling AI agents. However, Jena cautioned against using Microsoft’s benchmark results to predict enterprise savings. He said organizations should measure “cost and latency per successfully completed workflow, including fallbacks and mistakes.” Governance risks Decision-1 returns scores but does not provide a rationale for each decision. Microsoft’s documentation also warns that results can change depending on how questions are worded, Jena said. “My concern is that a precise-looking probability can give an uncertain decision more authority than it deserves,” he said. Dai said the absence of reasoning can make decisions harder to explain, validate, and challenge. He recommended calibration testing, controls over question wording and score thresholds, audit logging, and human review for high-impact decisions, particularly when customer outcomes or regulatory obligations are involved. Jena recommended beginning with reversible tasks, such as request routing and triage, and requiring meaningful human review for decisions that materially affect people. The concerns center on how organizations interpret the model’s scores and use them to trigger actions. A probability assigned to an answer does not, by itself, explain the decision or establish that it is correct for a particular enterprise application. The article originally appeared on InfoWorld.