The tools are available to everyone. The subscription is company-wide. The training sessions have been held. And yet, in offices from Wall Street to Silicon Valley, a stark divide is opening between workers who have woven artificial intelligence into the fabric of their daily work and colleagues who have barely touched it.The gap is not small. According to a <a href="https://cdn.openai.com/pdf/7ef17d82-96bf-4dd1-9df2-228f7f377a29/the-state-of-enterprise-ai_2025-report.pdf">new report</a> from OpenAI analyzing usage patterns across its more than one million business customers, workers at the <a href="https://openai.com/index/the-state-of-enterprise-ai-2025-report/">95th percentile of AI adoption</a> are sending six times as many messages to ChatGPT as the median employee at the same companies. For specific tasks, the divide is even more dramatic: frontier workers send 17 times as many coding-related messages as their typical peers, and among data analysts, the heaviest users engage the data analysis tool 16 times more frequently than the median.This is not a story about access. It is a story about a new form of workplace stratification emerging in real time — one that may be reshaping who gets ahead, who falls behind, and what it means to be a skilled worker in the age of artificial intelligence.<h2>Everyone has the same tools, but not everyone is using them</h2>Perhaps the most striking finding in the <a href="https://cdn.openai.com/pdf/7ef17d82-96bf-4dd1-9df2-228f7f377a29/the-state-of-enterprise-ai_2025-report.pdf">OpenAI report</a> is how little access explains. <a href="https://chatgpt.com/business/enterprise?utm_source=google&amp;utm_medium=paidsearch_brand&amp;utm_campaign=GOOG_B_SEM_GBR_Core_ENT_BAU_ACQ_PER_BRD_ALL_NAMER_US_EN_080625&amp;utm_term=chatgpt%20enterprise&amp;utm_content=182507886919&amp;utm_ad=779434575256&amp;utm_match=b&amp;gad_source=1&amp;gad_campaignid=22855802308&amp;gbraid=0AAAAA-I0E5ew3OYocHAOSOiKw516uqob_&amp;gclid=Cj0KCQiArt_JBhCTARIsADQZaylE-8Q0a0VjiXwyRLvFkD9BIkgt4EpWCm6v7alzBtNKWKE5_7J_3SYaArnTEALw_wcB">ChatGPT Enterprise</a> is now deployed across more than 7 million workplace seats globally, a nine-fold increase from a year ago. The tools are the same for everyone. The capabilities are identical. And yet usage varies by orders of magnitude.Among monthly active users — people who have logged in at least once in the past 30 days — <a href="https://cdn.openai.com/pdf/7ef17d82-96bf-4dd1-9df2-228f7f377a29/the-state-of-enterprise-ai_2025-report.pdf">19 percent have never tried the data analysis feature</a>. Fourteen percent have never used reasoning capabilities. Twelve percent have never used search. These are not obscure features buried in submenus; they are core functionality that OpenAI highlights as transformative for knowledge work.The pattern inverts among daily users. <a href="https://cdn.openai.com/pdf/7ef17d82-96bf-4dd1-9df2-228f7f377a29/the-state-of-enterprise-ai_2025-report.pdf">Only 3 percent</a> of people who use ChatGPT every day have never tried data analysis; just 1 percent have skipped reasoning or search. The implication is clear: the divide is not between those who have access and those who don&#x27;t, but between those who have made AI a daily habit and those for whom it remains an occasional novelty.<h2>Employees who experiment more are saving dramatically more time</h2>The <a href="https://cdn.openai.com/pdf/7ef17d82-96bf-4dd1-9df2-228f7f377a29/the-state-of-enterprise-ai_2025-report.pdf">OpenAI report</a> suggests that AI productivity gains are not evenly distributed across all users but concentrated among those who use the technology most intensively. Workers who engage across approximately seven distinct task types — data analysis, coding, image generation, translation, writing, and others — report saving five times as much time as those who use only four. Employees who save more than 10 hours per week consume eight times more AI credits than those who report no time savings at all.This creates a compounding dynamic. Workers who experiment broadly discover more uses. More uses lead to greater productivity gains. Greater productivity gains presumably lead to better performance reviews, more interesting assignments, and faster advancement—which in turn provides more opportunity and incentive to deepen AI usage further.Seventy-five percent of surveyed workers report being able to complete tasks they previously could not perform, including programming support, spreadsheet automation, and technical troubleshooting. For workers who have embraced these capabilities, the boundaries of their roles are expanding. For those who have not, the boundaries may be contracting by comparison.<h2>The corporate AI paradox: $40 billion spent, 95 percent seeing no return</h2>The individual usage gap documented by OpenAI mirrors a broader pattern identified by a separate study from <a href="https://venturebeat.com/ai/mit-report-misunderstood-shadow-ai-economy-booms-while-headlines-cry-failure">MIT&#x27;s Project NANDA</a>. Despite $30 billion to $40 billion invested in generative AI initiatives, only 5 percent of organizations are seeing transformative returns. The researchers call this the &quot;<a href="https://www.searchyour.ai/en/genai-divide-state-ai-business-2025-mit-nanda">GenAI Divide</a>&quot; — a gap separating the few organizations that succeed in transforming processes with adaptive AI systems from the majority that remain stuck in pilots.The MIT report found <a href="https://www.legal.io/articles/5719519/MIT-Report-Finds-95-of-AI-Pilots-Fail-to-Deliver-ROI-Exposing-GenAI-Divide">limited disruption</a> across industries: only two of nine major sectors—technology and media—show material business transformation from generative AI use. Large firms lead in pilot volume but lag in successful deployment.The pattern is consistent across both studies. Organizations and individuals are buying the technology. They are launching pilots. They are attending training sessions. But somewhere between adoption and transformation, most are getting stuck.<h2>While official AI projects stall, a shadow economy is thriving</h2>The <a href="https://venturebeat.com/ai/mit-report-misunderstood-shadow-ai-economy-booms-while-headlines-cry-failure">MIT study</a> reveals a striking disconnect: while only 40 percent of companies have purchased official LLM subscriptions, employees in over 90 percent of companies regularly use personal AI tools for work. Nearly every respondent reported using LLMs in some form as part of their regular workflow.&quot;This &#x27;<a href="https://venturebeat.com/ai/mit-report-misunderstood-shadow-ai-economy-booms-while-headlines-cry-failure">shadow AI</a>&#x27; often delivers better ROI than formal initiatives and reveals what actually works for bridging the divide,&quot; MIT&#x27;s Project NANDA found.The shadow economy offers a clue to what&#x27;s happening at the individual level within organizations. Employees who take initiative — who sign up for personal subscriptions, who experiment on their own time, who figure out how to integrate AI into their workflows without waiting for IT approval — are pulling ahead of colleagues who wait for official guidance that may never come.These shadow systems, largely unsanctioned, often deliver better performance and faster adoption than corporate tools. Worker sentiment reveals a preference for flexible, responsive tools — precisely the kind of experimentation that separates OpenAI&#x27;s frontier workers from the median.<h2>The biggest gaps show up in technical work that used to require specialists</h2>The largest relative gaps between frontier and median workers appear in coding, writing, and analysis — precisely the task categories where AI capabilities have advanced most rapidly. Frontier workers are not just doing the same work faster; they appear to be doing different work entirely, expanding into technical domains that were previously inaccessible to them.Among <a href="https://chatgpt.com/business/enterprise?utm_source=google&amp;utm_medium=paidsearch_brand&amp;utm_campaign=GOOG_B_SEM_GBR_Core_ENT_BAU_ACQ_PER_BRD_ALL_NAMER_US_EN_080625&amp;utm_term=chatgpt%20enterprise&amp;utm_content=182507886919&amp;utm_ad=779434575253&amp;utm_match=b&amp;gad_source=1&amp;gad_campaignid=22855802308&amp;gbraid=0AAAAA-I0E5ew3OYocHAOSOiKw516uqob_&amp;gclid=Cj0KCQiArt_JBhCTARIsADQZaykhZ9zJ10fojuPYA8XhHxR0jVT-WbgS4kfA7IIaTohf7yAByZszI-AaAkeREALw_wcB">ChatGPT Enterprise</a> users outside of engineering, IT, and research, coding-related messages have grown 36 percent over the past six months. Someone in marketing or HR who learns to write scripts and automate workflows is becoming a categorically different employee than a peer who has not — even if they hold the same title and started with the same skills.The academic research on AI and productivity offers a complicated picture. Several studies cited in the OpenAI report find that AI has an &quot;<a href="https://cdn.openai.com/pdf/7ef17d82-96bf-4dd1-9df2-228f7f377a29/the-state-of-enterprise-ai_2025-report.pdf">equalizing effect</a>,&quot; disproportionately helping lower-performing workers close the gap with their higher-performing peers. But the equalizing effect may apply only within the population of workers who actually use AI regularly. A meaningful share of workers are not in that group at all. They remain light users or non-users, even as their more adventurous colleagues pull away.<h2>Companies are divided too, and the gap is widening by the month</h2>The divide is not only between individual workers. It exists between entire organizations.Frontier firms — those at the 95th percentile of adoption intensity — generate approximately twice as many AI messages per employee as the median enterprise. For messages routed through custom GPTs, purpose-built tools that automate specific workflows, the gap widens to seven-fold.These numbers suggest fundamentally different operating models. At median companies, AI may be a productivity tool that individual workers use at their discretion. At frontier firms, AI appears to be embedded in core infrastructure: standardized workflows, persistent custom tools, systematic integration with internal data systems.The <a href="https://cdn.openai.com/pdf/7ef17d82-96bf-4dd1-9df2-228f7f377a29/the-state-of-enterprise-ai_2025-report.pdf">OpenAI report</a> notes that roughly one in four enterprises still has not enabled connectors that give AI access to company data—a basic step that dramatically increases the technology&#x27;s utility. The MIT study found that companies that purchased AI tools from specialized vendors succeeded <a href="https://www.tekedia.com/the-genai-divide-mit-report-shows-why-most-business-ai-projects-are-stalling/">67 percent</a> of the time, while internal builds had only a one-in-three success rate. For many organizations, the AI era has technically arrived but has not yet begun in practice.<h2>The technology is no longer the problem — organizations are</h2>For executives, the data presents an uncomfortable challenge. The technology is no longer the constraint. OpenAI notes that it releases a new feature or capability roughly every three days; the models are advancing faster than most organizations can absorb. The bottleneck has shifted from what AI can do to whether organizations are structured to take advantage of it.&quot;The dividing line isn&#x27;t intelligence,&quot; the <a href="https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf">MIT authors write</a>. The problems with enterprise AI have to do with memory, adaptability, and learning capability. Problems stem less from regulations or model performance, and more from tools that fail to learn or adapt.Leading firms, according to the <a href="https://cdn.openai.com/pdf/7ef17d82-96bf-4dd1-9df2-228f7f377a29/the-state-of-enterprise-ai_2025-report.pdf">OpenAI report</a>, consistently invest in executive sponsorship, data readiness, workflow standardization, and deliberate change management. They build cultures where custom AI tools are created, shared, and refined across teams. They track performance and run evaluations. They make AI adoption a strategic priority rather than an individual choice.The rest are leaving it to chance — hoping that workers will discover the tools on their own, experiment on their own time, and somehow propagate best practices without infrastructure or incentive. The six-fold gap suggests this approach is not working.<h2>The window to catch up is closing faster than most companies realize</h2>With enterprise contracts locking in over the next 18 months, there&#x27;s a shrinking window for vendors and adopters to cross the divide.The GenAI Divide identified by the <a href="https://cdn.openai.com/pdf/7ef17d82-96bf-4dd1-9df2-228f7f377a29/the-state-of-enterprise-ai_2025-report.pdf">MIT report</a> is not going to last forever. But the organizations that figure out a way across it soonest will be the ones that define the next era of business.Both reports carry caveats. The OpenAI data comes from a company with an obvious interest in promoting AI adoption. The productivity figures are self-reported by customers already paying for the product. The MIT study, while independent, relies on interviews and surveys rather than direct measurement. The long-term effects of this technology on employment, wages, and workplace dynamics remain uncertain.But the core finding — that access alone does not produce adoption, and that adoption varies enormously even within organizations that have made identical tools available to all — is consistent with how previous technologies have diffused through the economy. Spreadsheets, email, and the internet all created similar divides before eventually becoming universal. The question is how long the current gap persists, who benefits during the transition, and what happens to workers who find themselves on the wrong side of it.For now, the divide is stark. Ninety percent of users said they prefer humans for &quot;mission-critical work,&quot; while AI has &quot;won the war for simple work.&quot; The workers who are pulling ahead are not doing so because they have access their colleagues lack. They are pulling ahead because they decided to use what everyone already has—and kept using it until they figured out what it could do.The 6x gap is not about technology. It is about behavior. And behavior, unlike software, cannot be deployed with a company-wide rollout.

يكشف تقرير حديث من OpenAI عن فجوة كبيرة في الإنتاجية، حيث يظهر أن مستخدمي الذكاء الاصطناعي المتقدمين يرسلون ستة أضعاف الرسائل إلى ChatGPT مقارنة بالموظف المتوسط في شركاتهم. هذه الفجوة أكثر وضوحًا في أدوار محددة، مثل البرمجة وتحليل البيانات، حيث يتفاعل المستخدمون الأكثر نشاطًا 17 مرة أكثر من نظرائهم.

Un informe reciente de OpenAI destaca una brecha significativa en la productividad, revelando que los usuarios avanzados de IA envían seis veces más mensajes a ChatGPT que el empleado medio en sus empresas. Esta disparidad es aún más pronunciada en roles específicos, como la programación y el análisis de datos, donde los usuarios más activos interactúan 17 veces más que sus compañeros.

Un rapport récent d'OpenAI met en lumière un écart de productivité significatif, révélant que les utilisateurs avancés de l'IA envoient six fois plus de messages à ChatGPT que l'employé médian dans leurs entreprises. Cette disparité est encore plus marquée dans des rôles spécifiques, tels que la programmation et l'analyse de données, où les meilleurs utilisateurs interagissent 17 fois plus que leurs pairs.

A recent report from OpenAI highlights a significant productivity gap, revealing that AI power users send six times more messages to ChatGPT than the median employee in their companies. This disparity is even more pronounced in specific roles, such as coding and data analysis, where top users engage 17 times more than their peers.

OpenAI report reveals a 6x productivity gap between AI power users and everyone else

A San Francisco-based startup has demonstrated what it calls a breakthrough in hardware development: an artificial intelligence system that designed a fully functional Linux computer in one week — a process that would typically consume nearly three months of skilled engineering labor.<a href="https://www.quilter.ai/">Quilter</a>, which has raised more than $40 million from investors including Benchmark, Index Ventures, and Coatue, used its physics-driven AI to automate the design of a two-board computer system that booted successfully on its first attempt, requiring no costly revisions. The project, internally dubbed &quot;<a href="https://www.quilter.ai/blog">Project Speedrun</a>,&quot; required just 38.5 hours of human labor compared to the 428 hours that professional PCB designers quoted for the same task.The announcement also marks the first public disclosure that <a href="https://www.buildc.com/">Tony Fadell</a>, the engineer who led development of the iPod and iPhone at Apple and later founded Nest, has invested in the company and serves as an advisor.&quot;We didn&#x27;t teach Quilter to draw; we taught it to think in physics,&quot; said Sergiy Nesterenko, Quilter&#x27;s chief executive and a former SpaceX engineer, in an exclusive interview with VentureBeat. &quot;The result wasn&#x27;t a simulation — it was a working computer.&quot;<h2>Circuit board design remains the forgotten bottleneck that delays nearly every hardware product</h2>The announcement shines a light on an unglamorous but critical chokepoint in technology development: printed circuit board layout. While semiconductors and software have received enormous attention and investment, the green fiberglass boards that connect chips, memory, and components in virtually every electronic device remain stubbornly manual to design.&quot;Besides auto-routers, the technology really hadn&#x27;t changed since the early &#x27;90s,&quot; Fadell told VentureBeat. &quot;The best boards are still made by hand. You go to Apple, they&#x27;ve got the tools, and these guys are just pushing traces, checking everything, doing flood fills—and you&#x27;re like, there&#x27;s got to be a better way.&quot;The PCB design process typically unfolds in three stages. Engineers first create a schematic — a logical diagram showing how components connect. Then a specialist manually draws the physical layout in CAD software, placing components and routing thousands of copper traces across multiple layers. Finally, the design goes to a manufacturer for fabrication.That middle step — the layout — creates a persistent bottleneck. For a board of moderate complexity, the process typically consumes four to eight weeks. For sophisticated systems like computers or automotive electronics, timelines stretch to three months or longer.&quot;The timeline was always this elastic thing—they&#x27;d say, &#x27;Yeah, that&#x27;s two weeks minimum,&#x27;&quot; Fadell recalled of his experience at Apple and Nest. &quot;And we&#x27;d say, &#x27;No, no. Work day and night. It&#x27;s two weeks.&#x27; But it was always this fixed bottleneck.&quot;The consequences ripple through hardware organizations. Firmware teams sit idle waiting for physical boards to test their code. Validation engineers cannot begin debugging. Product launches slip. According to Quilter&#x27;s research, only about 10 percent of first board revisions work correctly, forcing expensive and time-consuming respins.<h2>Project Speedrun put Quilter&#x27;s AI to the test with an 843-component computer that booted on the first try</h2><a href="https://www.quilter.ai/blog">Project Speedrun</a> was designed to push the technology to its limits while producing an easily understood result: a working computer that could boot Linux, browse the internet, and run applications.The system consists of two boards based on<a href="http://i.mx"> NXP&#x27;s i.MX 8M Mini</a> reference platform, a processor architecture used in automotive infotainment, industrial automation, and machine vision applications.The main system-on-module contains a quad-core ARM processor running at 1.8 gigahertz, 2 gigabytes of LPDDR4 memory, and 32 gigabytes of eMMC storage. A companion baseboard provides connectivity including Ethernet, USB, HDMI, and audio.Together, the boards incorporate 843 components and 5,141 electrical connections, or &quot;pins,&quot; routed across eight-layer circuit board stackups manufactured by Sierra Circuits in California. The minimum trace geometry reached 2 mils (two-thousandths of an inch) on the system-on-module — fine enough to require advanced high-density interconnect manufacturing techniques.Quilter&#x27;s AI completed the layout with approximately 98 percent routing coverage and zero design rule violations. Both boards passed power-on testing and successfully booted <a href="https://www.debian.org/">Debian Linux</a> on the first attempt.&quot;We made an entire computer to demonstrate that this technology works,&quot; Nesterenko said. &quot;We took something that&#x27;s typically quoted at 400 to 450 hours, automated the vast majority of it, and reduced it to about 30 to 40 hours of cleanup time.&quot;The cleanup time is work that human engineers still perform: reviewing the AI&#x27;s output, fixing any issues, and preparing final fabrication files. But even with that overhead, the total elapsed time from schematic to fabricated boards collapsed from the typical 11 weeks to a single week.<h2>Unlike ChatGPT, Quilter&#x27;s AI learns by playing billions of games against the laws of physics</h2>Quilter&#x27;s <a href="https://www.quilter.ai/product">technical approach</a> differs fundamentally from the large language models that have dominated recent AI headlines. Where systems like GPT-5 or Claude learn to predict text based on massive training datasets of human writing, <a href="https://www.quilter.ai/product/technology">Quilter&#x27;s AI</a> learns by playing what amounts to an elaborate game against the laws of physics.&quot;Language models don&#x27;t apply to us because this is not a language problem,&quot; Nesterenko explained. &quot;If you ask it to actually create a blueprint, it has no training data for that. It has no context for that.&quot;The company also rejected the seemingly obvious approach of training on examples of human-designed boards. Nesterenko cited three reasons: humans make frequent errors (explaining why most boards require revisions), the best designs are locked inside large companies unwilling to share proprietary data, and training on human examples would cap the AI&#x27;s performance at human levels.Instead, Quilter built what Nesterenko describes as a &quot;game&quot; where the AI agent makes sequential decisions — place this component here, route this trace there — and receives feedback based on whether the resulting design satisfies electromagnetic, thermal, and manufacturing constraints.&quot;What you&#x27;re really changing is not the probability of getting a very specific outcome of the model, but the probability of choosing a certain action based on that experience,&quot; Nesterenko said.The approach mirrors DeepMind&#x27;s progression with its Go-playing systems. The original <a href="https://deepmind.google/research/alphago/">AlphaGo</a> learned from human games, but its successor <a href="https://deepmind.google/research/alphazero-and-muzero/">AlphaZero</a> learned purely through self-play and ultimately surpassed human capability. Quilter harbors similar ambitions.&quot;In the long term, to come up with better designs for circuit boards than humans have ever tried to do,&quot; Nesterenko said.Fadell drew a parallel to an earlier technological transition: &quot;I remember this with assembly. You had assembly and compilers, and engineers would say, &#x27;I can&#x27;t trust the compiler. I&#x27;m going to do the loop unrolling myself.&#x27; Now very, very few people write any assembly.&quot;He expects PCB design to follow a similar arc: &quot;I hope the same thing happens with PCB design. Sure, a few people will hold out, but these tools are going to get so good that everyone else will move on.&quot;<h2>Fadell and Nesterenko spent months solving a delicate problem: how to automate design without stripping engineers of control</h2>Automating a task that skilled professionals have performed manually for decades raises an obvious question: how do engineers maintain control over designs that will ultimately ship in products where reliability matters?Fadell said he spent significant time with Nesterenko working through this tension. The solution, he said, lies in allowing users to choose their level of involvement at each stage of the process.&quot;If you&#x27;re a control freak, you can be a control freak. If you want to say &#x27;just do it for me,&#x27; you can do that too—and everything in between,&quot; Fadell said. &quot;You can walk through each phase of the design and get involved wherever you want, or let the AI handle it.&quot;The workflow breaks into three phases: setup, where engineers define constraints and requirements; execution, where the AI generates candidate layouts; and cleanup, where humans review and refine the output. Engineers can intervene at any point, adjusting constraints and regenerating designs until they&#x27;re satisfied.&quot;This is something Tony and I talk about a lot,&quot; Nesterenko said. &quot;How do we give users control while still automating most of the work?&quot;<h2>Quilter&#x27;s technology has clear boundaries: 10,000 pins and 10 gigahertz mark the current limits</h2>The technology has clear limitations. <a href="https://www.quilter.ai/">Quilter</a> currently handles boards with up to roughly 10,000 pins — sufficient for a wide range of applications but well short of the most complex designs, which can exceed 100,000 connections.Physics complexity also creates boundaries. The system handles high-speed communications up to approximately 10 gigahertz, covering typical consumer electronics and many industrial applications. But advanced systems like sophisticated radar, which can operate at 100 gigahertz, exceed current capabilities.&quot;There are boards where Quilter won&#x27;t make enough progress to make the cleanup time worthwhile,&quot; Nesterenko acknowledged. &quot;We&#x27;re just not that helpful yet with the most advanced, sophisticated designs.&quot;The company has focused initially on categories where speed matters more than extreme complexity: test fixtures, evaluation boards, design validation boards, and environmental test hardware. These boards often sit in long queues behind higher-priority production designs, delaying engineering programs.<h2>The company bets that engineers will pay the same price for a 10x speed improvement</h2>Quilter <a href="https://www.quilter.ai/pricing">prices its service by pin count</a>, matching the billing conventions that already exist when companies hire outside layout specialists. The pitch to customers is cost neutrality with a ten-fold improvement in speed.&quot;We&#x27;re going to charge you roughly the same that you would pay for the pins that you would with a person,&quot; Nesterenko said. &quot;But the reason you choose us is that we do this 10 times faster.&quot;For a company waiting three months for a board layout, receiving it in a week fundamentally changes what&#x27;s possible. Engineering teams can run multiple design experiments in parallel. Firmware developers get hardware faster. Products reach the market sooner.The company offers free access for hobbyists, students, and small businesses with less than $50,000 in revenue — a strategy to build familiarity while targeting enterprise customers for commercial revenue.<h2>The iPod creator waited years to attach his name to Quilter — until he could prove the technology actually works</h2>Fadell said he chose this moment to publicly acknowledge his investment because the <a href="https://www.quilter.ai/blog">Project Speedrun</a> demonstration provides concrete evidence that the technology works.&quot;It&#x27;s not about being comfortable—I was always comfortable with the team,&quot; he said. &quot;This was about waiting until we had something you could hang your hat on. Now I can say, &#x27;I&#x27;ve used the tool. I&#x27;ve seen it.&#x27;&quot;He contrasted his approach with typical investor announcements: &quot;Every investor goes, I invested in this, it&#x27;s gonna change the world. It&#x27;s like, no, I know better. I&#x27;ve used the tool. I know people who use it. I asked my startups to use the tool.&quot;Fadell&#x27;s involvement goes beyond capital. He described email exchanges running to &quot;a dozen pages of details&quot; covering product design, user experience, enterprise sales, and technical architecture.&quot;Of all the investors I work with, Tony by far goes deepest with me on the product side,&quot; Nesterenko said.<h2>If Quilter succeeds, it could unlock a new generation of hardware startups that were never economically viable before</h2>The stakes extend far beyond one company&#x27;s product roadmap. If Quilter&#x27;s technology scales, it could fundamentally alter the economics of building physical products.Fadell argued that hardware development has historically moved slowly because each step in the process — schematic design, PCB layout, manufacturing, assembly — created friction. Other innovations have already smoothed schematic tools and manufacturing. Layout remained the stubborn holdout.&quot;Once you shrink that from weeks to hours, you can iterate so much faster because all the other friction in the chain has been reduced,&quot; Fadell said.He predicted the technology would eventually extend upstream into schematic design itself, with AI that understands both logical connections and physical constraints helping engineers avoid problems earlier in the process.At MIT, where Fadell now spends time, he encounters would-be founders who have abandoned hardware ambitions because the process seemed insurmountable.&quot;I talk to professors and startup founders, and they say, &#x27;I&#x27;m never doing hardware. It&#x27;s too hard,&#x27;&quot; he said. &quot;I hope we can make it easier for more people to jump in and try things.&quot;Industry veterans remain skeptical. Auto-routing tools — previous attempts at automation — became notorious for producing unusable results, spawning T-shirts proclaiming engineers would &quot;<a href="https://www.eevblog.com/forum/chat/chris-gammel_s-_never-trust-the-autorouter-shirt_-on-teespring/">never trust the auto-router</a>.&quot;Nesterenko has seen the skepticism dissolve in real time. He described a recent meeting with executives from a major customer who came to discuss Quilter&#x27;s capabilities. As the conversation unfolded, one executive picked up the <a href="https://www.quilter.ai/blog">Project Speedrun</a> boards and began photographing them from every angle, turning them over in his hands.&quot;He was just fascinated by the fact that this is possible now,&quot; Nesterenko said.The question is no longer whether AI can design circuit boards. A working Linux computer, assembled from 843 components and booted on the first attempt, answers that definitively. The question now is what engineers will build when layout stops being the bottleneck — when hardware, as Fadell put it, finally &quot;moves at the speed of thought.&quot;On that point, Nesterenko offered a prediction. &quot;If you ask the average electrical engineer today whether automation or AI could at all help with the board of this complexity, they would say no,&quot; he said. For decades, they would have been right. As of last week, they&#x27;re not.

حققت شركة كويلتر، وهي شركة ناشئة مقرها سان فرانسيسكو، إنجازًا كبيرًا من خلال استخدام الذكاء الاصطناعي لتصميم جهاز كمبيوتر يعمل بنظام لينكس في غضون أسبوع واحد فقط، وهي عملية تستغرق عادةً ما يقرب من ثلاثة أشهر من العمل الهندسي الماهر. المشروع، الذي أطلق عليه اسم Project Speedrun، نتج عنه نظام كمبيوتر مكون من لوحين بدأ بنجاح من المحاولة الأولى، مما تطلب فقط 38.5 ساعة من العمل البشري مقارنةً بـ 428 ساعة التي قدرها مصممو الدوائر المحترفون.

Una startup con sede en San Francisco, Quilter, ha logrado un hito significativo al utilizar inteligencia artificial para diseñar una computadora Linux completamente funcional en solo una semana, una tarea que normalmente lleva casi tres meses a ingenieros calificados. El proyecto, denominado Project Speedrun, resultó en un sistema de computadora de dos placas que se inició con éxito en su primer intento, requiriendo solo 38.5 horas de trabajo humano en comparación con las 428 horas estimadas por diseñadores profesionales.

Une startup basée à San Francisco, Quilter, a réalisé une avancée majeure en utilisant l'intelligence artificielle pour concevoir un ordinateur Linux entièrement fonctionnel en seulement une semaine, une tâche qui prend généralement près de trois mois aux ingénieurs qualifiés. Le projet, nommé Project Speedrun, a abouti à un système informatique à deux cartes qui a réussi à démarrer dès sa première tentative, nécessitant seulement 38,5 heures de travail humain contre 428 heures estimées par des concepteurs professionnels.

A San Francisco-based startup, Quilter, has achieved a significant milestone by using artificial intelligence to design a fully functional Linux computer in just one week, a task that typically takes skilled engineers nearly three months. The AI-driven project, named Project Speedrun, resulted in a two-board computer system that successfully booted on its first attempt, requiring only 38.5 hours of human labor compared to the 428 hours estimated by professional designers.

Quilter's AI just designed an 843‑part Linux computer that booted on the first try. Hardware will never be the same.

French AI startup Mistral has weathered a rocky period of public questioning over the last year to emerge, now here in December 2025, with new, crowd-pleasing models for enterprise and indie developers.Just days after releasing its <a href="https://venturebeat.com/ai/mistral-launches-mistral-3-a-family-of-open-models-designed-to-run-on">powerful open source, general purpose Mistral 3 LLM family</a> for edge devices and local hardware, the <a href="https://mistral.ai/news/devstral-2-vibe-cli">company returned today to debut Devstral 2</a>.The release includes a new pair of models optimized for software engineering tasks — again, with one small enough to run on a single laptop, offline and privately — alongside Mistral Vibe, a command-line interface (CLI) agent designed to allow developers to call the models up directly within their terminal environments. The models are fast, lean, and open—at least in theory. But the real story lies not just in the benchmarks, but in how Mistral is packaging this capability: one model fully free, another conditionally so, and a terminal interface built to scale with either.It’s an attempt not just to match proprietary systems like Claude and GPT-4 in performance, but to compete with them on developer experience—and to do so while holding onto the flag of open-source.Both models are available now for free for a limited time <a href="https://docs.mistral.ai/models/devstral-2-25-12">via Mistral’s API</a> and <a href="https://huggingface.co/collections/mistralai/devstral-2">Hugging Face</a>. The full Devstral 2 model is supported out-of-the-box in the community inference provider <a href="https://x.com/vllm_project/status/1998428798891765926?s=20">vLLM</a> and on the open source agentic coding platform <a href="https://x.com/kilocode/status/1998412042357588461">Kilo Code</a>. <h2>A Coding Model Meant to Drive</h2>At the top of the announcement is Devstral 2, a 123-billion parameter dense transformer with a 256K-token context window, engineered specifically for agentic software development. Mistral says the model achieves 72.2% on SWE-bench Verified, a benchmark designed to evaluate long-context software engineering tasks in real-world repositories.The smaller sibling, Devstral Small 2, weighs in at 24B parameters, with the same long context window and a performance of 68.0% on SWE-bench. On paper, that makes it the strongest open-weight model of its size, even outscoring many 70B-class competitors.But the performance story isn’t just about raw percentages. Mistral is betting that efficient intelligence beats scale, and has made much of the fact that Devstral 2 is:<ul><li>5× smaller than DeepSeek V3.2</li><li>8× smaller than Kimi K2</li><li>Yet still matches or surpasses them on key software reasoning benchmarks.</li></ul>Human evaluations back this up. In side-by-side comparisons:<ul><li>Devstral 2 beat DeepSeek V3.2 in 42.8% of tasks, losing only 28.6%.</li><li>Against Claude Sonnet 4.5, it lost more often (53.1%)—a reminder that while the gap is narrowing, closed models still lead in overall preference.</li></ul>Still, for an open-weight model, these results place Devstral 2 at the frontier of what’s currently available to run and modify independently.<h2>Vibe CLI: A Terminal-Native Agent</h2>Alongside the models, Mistral released Vibe CLI, a command-line assistant that integrates directly with Devstral models. It’s not an IDE plugin or a ChatGPT-style code explainer. It’s a native interface designed for project-wide code understanding and orchestration, built to live inside the developer’s actual workflow.Vibe brings a surprising degree of intelligence to the terminal:<ul><li>It reads your file tree and Git status to understand project scope.</li><li>It lets you reference files with @, run shell commands with !, and toggle behavior with slash commands.</li><li>It orchestrates changes across multiple files, tracks dependencies, retries failed executions, and can even refactor at architectural scale.</li></ul>Unlike most developer agents, which simulate a REPL from within a chat UI, Vibe starts with the shell and pulls intelligence in from there. It’s programmable, scriptable, and themeable. And it’s released under the Apache 2.0 license, meaning it’s truly free to use—in commercial settings, internal tools, or open-source extensions.<h2>Licensing Structure: Open-ish — With Revenue Limitations</h2>At first glance, Mistral’s licensing approach appears straightforward: the models are open-weight and publicly available. But a closer look reveals a line drawn through the middle of the release, with different rules for different users.Devstral Small 2, the 24-billion parameter variant, is covered under a standard, enterprise- and developer-friendly <a href="https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/apache-2.0.md">Apache 2.0 license</a>. That’s a gold standard in open-source: no revenue restrictions, no fine print, no need to check with legal. Enterprises can use it in production, embed it into products, and redistribute fine-tuned versions without asking for permission.Devstral 2, the flagship 123B model, is released under what Mistral calls a “<a href="https://huggingface.co/mistralai/Devstral-2-123B-Instruct-2512/blob/main/LICENSE">modified MIT license</a>.” That phrase sounds innocuous, but the modification introduces a critical limitation: any company making more than $20 million in monthly revenue cannot use the model at all—not even internally—without securing a separate commercial license from Mistral.<blockquote>“You are not authorized to exercise any rights under this license if the global consolidated monthly revenue of your company […] exceeds $20 million,” the license reads.</blockquote>The clause applies not only to the base model, but to derivatives, fine-tuned versions, and redistributed variants, regardless of who hosts them. In effect, it means that while the weights are “open,” their use is gated for large enterprises—unless they’re willing to engage with Mistral’s sales team or use the hosted API at metered pricing.To draw an analogy: Apache 2.0 is like a public library—you walk in, borrow the book, and use it however you need. Mistral’s modified MIT license is more like a corporate co-working space that’s free for freelancers but charges rent once your company hits a certain size.<h2>Weighing Devstral Small 2 for Enterprise Use</h2>This division raises an obvious question for larger companies: can Devstral Small 2 with its more permissive and unrestricted Apache 2.0 licensing serve as a viable alternative for medium-to-large enterprises?The answer depends on context. Devstral Small 2 scores 68.0% on SWE-bench, significantly ahead of many larger open models, and remains deployable on single-GPU or CPU-only setups. For teams focused on:<ul><li>internal tooling,</li><li>on-prem deployment,</li><li>low-latency edge inference,…it offers a rare combination of legality, performance, and convenience.</li></ul>But the performance gap from Devstral 2 is real. For multi-agent setups, deep monorepo refactoring, or long-context code analysis, that 4-point benchmark delta may understate the actual experience difference.For most enterprises, Devstral Small 2 will serve either as a low-friction way to prototype—or as a pragmatic bridge until licensing for Devstral 2 becomes feasible. It is not a drop-in replacement for the flagship, but it may be “good enough” in specific production slices, particularly when paired with Vibe CLI.But because Devstral Small 2 can be run entirely offline — including on a single GPU machine or a sufficiently specced laptop — it unlocks a critical use case for developers and teams operating in tightly controlled environments. Whether you’re a solo indie building tools on the go, or part of a company with strict data governance or compliance mandates, the ability to run a performant, long-context coding model without ever hitting the internet is a powerful differentiator. No cloud calls, no third-party telemetry, no risk of data leakage — just local inference with full visibility and control.This matters in industries like finance, healthcare, defense, and advanced manufacturing, where data often cannot leave the network perimeter. But it’s just as useful for developers who prefer autonomy over vendor lock-in — or who want their tools to work the same on a plane, in the field, or inside an air-gapped lab. In a market where most top-tier code models are delivered as API-only SaaS products, Devstral Small 2 offers a rare level of portability, privacy, and ownership.In that sense, Mistral isn’t just offering open models—they’re offering multiple paths to adoption, depending on your scale, compliance posture, and willingness to engage.<h2>Integration, Infrastructure, and Access</h2>From a technical standpoint, Mistral’s models are built for deployment. Devstral 2 requires a minimum of 4× H100-class GPUs, and is already available on build.nvidia.com. Devstral Small 2 can run on a single GPU or CPU such as those in a standard laptop, making it accessible to solo developers and embedded teams alike.Both models support quantized FP4 and FP8 weights, and are compatible with vLLM for scalable inference. Fine-tuning is supported out of the box.API pricing—after the free introductory window—follows a token-based structure:<ul><li>Devstral 2: $0.40 per million input tokens / $2.00 for output</li><li>Devstral Small 2: $0.10 input / $0.30 output</li></ul>That pricing sits just below OpenAI’s GPT-4 Turbo, and well below Anthropic’s Claude Sonnet at comparable performance levels.<h2>Developer Reception: Ground-Level Buzz</h2>On X (formerly Twitter), developers reacted quickly with a wave of positive reception, with Hugging Face&#x27;s Head of Product <a href="https://x.com/victormustar/status/1998414127400923246">Victor Mustar asking</a> if the small, Apache 2.0 licensed variant was the &quot;new local coding king,&quot; i.e., the one developers could use to run on their laptops directly and privately, without an internet connection:<div></div>Another popular AI news and rumors account, TestingCatalogNews, posted that it was &quot;SOTTA in coding,&quot; or &quot;State Of The Tiny Art&quot;<div></div>Another user, <a href="https://x.com/xlr8harder/status/1998458990565396505">@xlr8harder</a>, took issue with the custom licensing terms for Devstral 2, writing &quot;calling the Devstral 2 license &#x27;modified MIT&#x27; is misleading at best. It’s a proprietary license with MIT-like attribution requirements.&quot;<div></div>While the tone was critical, it reflected some attention Mistral’s license structuring was receiving, particularly among developers familiar with open-use norms.<h2>Strategic Context: From Codestral to Devstral and Mistral 3</h2>Mistral’s steady push into software development tools didn’t start with Devstral 2—it began in May 2024 with <a href="https://venturebeat.com/ai/mistral-announces-codestral-its-first-programming-focused-ai-model">Codestral</a>, the company’s first code-focused large language model. A 22-billion parameter system trained on more than 80 programming languages, Codestral was designed for use in developer environments ranging from basic autocompletions to full function generation. The model launched under a non-commercial license but still outperformed heavyweight competitors like CodeLlama 70B and Deepseek Coder 33B in early benchmarks such as HumanEval and RepoBench.Codestral’s release marked Mistral’s first move into the competitive coding-model space, but it also established a now-familiar pattern: technically lean models with surprisingly strong results, a wide context window, and licensing choices that invited developer experimentation. Industry partners including JetBrains, LlamaIndex, and LangChain quickly began integrating the model into their workflows, citing its speed and tool compatibility as key differentiators.One year later, the company followed up with <a href="https://venturebeat.com/ai/mistral-ai-launches-devstral-powerful-new-open-source-swe-agent-model-that-runs-on-laptops">Devstral</a>, a 24B model purpose-built for “agentic” behavior—handling long-range reasoning, file navigation, and autonomous code modification. Released in partnership with All Hands AI and licensed under Apache 2.0, Devstral was notable not just for its portability (it could run on a MacBook or RTX 4090), but for its performance: it beat out several closed models on SWE-Bench Verified, a benchmark of 500 real-world GitHub issues.Then came Mistral 3, announced in December 2025 as <a href="https://venturebeat.com/ai/mistral-launches-mistral-3-a-family-of-open-models-designed-to-run-on">a portfolio of 10 open-weight models</a> targeting everything from drones and smartphones to cloud infrastructure. This suite included both high-end models like Mistral Large 3 (a MoE system with 41 active parameters and 256K context) and lightweight “Ministral” variants that could run on 4GB of VRAM. All were licensed under Apache 2.0, reinforcing Mistral’s commitment to flexible, edge-friendly deployment.Mistral 3 positioned the company not as a direct competitor to frontier models like GPT-5 or Gemini 3, but as a developer-first platform for customized, localized AI systems. Co-founder Guillaume Lample described the vision as “distributed intelligence”—many smaller systems tuned for specific tasks and running outside centralized infrastructure. “In more than 90% of cases, a small model can do the job,” he told VentureBeat. “It doesn’t have to be a model with hundreds of billions of parameters.”That broader strategy helps explain the significance of Devstral 2. It’s not a one-off release but a continuation of Mistral’s long-running commitment to code agents, local-first deployment, and open-weight availability—an ecosystem that began with Codestral, matured through Devstral, and scaled up with Mistral 3. Devstral 2, in this framing, is not just a model. It’s the next version of a playbook that’s been unfolding in public for over a year.<h2>Final Thoughts (For Now): A Fork in the Road</h2>With Devstral 2, Devstral Small 2, and Vibe CLI, Mistral AI has drawn a clear map for developers and companies alike. The tools are fast, capable, and thoughtfully integrated. But they also present a choice—not just in architecture, but in how and where you’re allowed to use them.If you’re an individual developer, small startup, or open-source maintainer, this is one of the most powerful AI systems you can freely run today. If you’re a Fortune 500 engineering lead, you’ll need to either talk to Mistral—or settle for the smaller model and make it work.In a market increasingly dominated by black-box models and SaaS lock-ins, Mistral’s offer is still a breath of fresh air. Just read the fine print before you start building.

أطلقت شركة ميسترال الفرنسية الناشئة في مجال الذكاء الاصطناعي نموذج البرمجة Devstral 2، الذي يتضمن إصدارًا مناسبًا لأجهزة الكمبيوتر المحمولة، مُحسّنًا لمهام هندسة البرمجيات. يأتي هذا الإطلاق بعد تقديم عائلة نماذج Mistral 3، التي تهدف إلى تحسين قدرات الأجهزة المحلية للمطورين.

La startup francesa de IA Mistral ha lanzado el modelo de codificación Devstral 2, que incluye una versión amigable para laptops optimizada para tareas de ingeniería de software. Este lanzamiento sigue a la introducción de la familia de modelos Mistral 3, destinada a mejorar las capacidades del hardware local para desarrolladores.

La startup française d'IA Mistral a lancé le modèle de codage Devstral 2, qui comprend une version adaptée aux ordinateurs portables, optimisée pour les tâches d'ingénierie logicielle. Ce lancement fait suite à l'introduction de la famille de modèles Mistral 3, visant à améliorer les capacités du matériel local pour les développeurs.

French AI startup Mistral has launched the Devstral 2 coding model, which includes a laptop-friendly version optimized for software engineering tasks. This release follows the introduction of the Mistral 3 LLM family, aimed at enhancing local hardware capabilities for developers.

Mistral launches powerful Devstral 2 coding model including open source, laptop-friendly version

There is no shortage of AI benchmarks in the market today, with popular options like<a href="https://venturebeat.com/ai/beyond-arc-agi-gaia-and-the-search-for-a-real-intelligence-benchmark"> Humanity&#x27;s Last Exam</a> (HLE), ARC-AGI-2 and GDPval, among numerous others.AI agents excel at solving abstract math problems and passing PhD-level exams that most benchmarks are based on, but Databricks has a question for the enterprise: Can they actually handle the document-heavy work most enterprises need them to do?The answer, according to new research from the data and AI platform company, is sobering. Even the best-performing AI agents achieve less than 45% accuracy on tasks that mirror real enterprise workloads, exposing a critical gap between academic benchmarks and business reality.&quot;If we focus our research efforts on getting better at [existing benchmarks], then we&#x27;re probably not solving the right problems to make Databricks a better platform,&quot; Erich Elsen, principal research scientist at Databricks, explained to VentureBeat. &quot;So that&#x27;s why we were looking around. How do we create a benchmark that, if we get better at it, we&#x27;re actually getting better at solving the problems that our customers have?&quot;The result is OfficeQA, a benchmark designed to test AI agents on grounded reasoning: Answering questions based on complex proprietary datasets containing unstructured documents and tabular data. Unlike existing benchmarks that focus on abstract capabilities, OfficeQA proxies for the economically valuable tasks enterprises actually perform.<h2>Why academic benchmarks miss the enterprise mark</h2>There are numerous shortcomings of popular AI benchmarks from an enterprise perspective, according to Elsen. HLE features questions requiring PhD-level expertise across diverse fields. ARC-AGI evaluates abstract reasoning through visual manipulation of colored grids. Both push the frontiers of AI capabilities, but don&#x27;t reflect daily enterprise work. Even GDPval, which was specifically created to evaluate economically useful tasks, misses the target.&quot;We come from a pretty heavy science or engineering background, and sometimes we create evals that reflect that,&quot; Elsen said. &quot; So they&#x27;re either extremely math-heavy, which is a great, useful task, but advancing the frontiers of human mathematics is not what customers are trying to do with Databricks.&quot;While AI is commonly used for customer support and coding apps, Databricks&#x27; customer base has a broader set of requirements. Elsen noted that answering questions about documents or corpora of documents is a common enterprise task. These require parsing complex tables with nested headers, retrieving information across dozens or hundreds of documents and performing calculations where a single-digit error can cascade into organizations making incorrect business decisions.<h2>Building a benchmark that mirrors enterprise document complexity</h2>To create a meaningful test of grounded reasoning capabilities, Databricks needed a dataset that approximates the messy reality of proprietary enterprise document corpora, while remaining freely available for research. The team landed on U.S. Treasury Bulletins, published monthly for five decades beginning in 1939 and quarterly thereafter.The Treasury Bulletins check every box for enterprise document complexity. Each bulletin runs 100 to 200 pages and consists of prose, complex tables, charts and figures describing Treasury operations: Where federal money came from, where it went and how it financed government operations. The corpus spans approximately 89,000 pages across eight decades. Until 1996, the bulletins were scans of physical documents; afterwards, they were digitally produced PDFs. USAFacts, an organization whose mission is &quot;to make government data easier to access and understand,&quot; partnered with Databricks to develop the benchmark, identifying Treasury Bulletins as ideal and ensuring questions reflected realistic use cases.The 246 questions require agents to handle messy, real-world document challenges: Scanned images, hierarchical table structures, temporal data spanning multiple reports and the need for external knowledge like inflation adjustments. Questions range from simple value lookups to multi-step analysis requiring statistical calculations and cross-year comparisons.To ensure the benchmark requires actual document-grounded retrieval, Databricks filtered out questions that LLMs could answer using parametric knowledge or web search alone. This removed simpler questions and some surprisingly complex ones where models leveraged historical financial records memorized during pre-training.Every question has a validated ground truth answer (typically a number, sometimes dates or small lists), enabling automated evaluation without human judging. This design choice matters: It allows reinforcement learning (RL) approaches that require verifiable rewards, similar to how models train on coding problems.<h2>Current performance exposes fundamental gaps</h2>Databricks tested Claude Opus 4.5 Agent (using Claude&#x27;s SDK) and GPT-5.1 Agent (using OpenAI&#x27;s File Search API). The results should give pause to any enterprise betting heavily on current agent capabilities.When provided with raw PDF documents:<ul><li> Claude Opus 4.5 Agent (with default thinking=high) achieved 37.4% accuracy. </li><li> GPT-5.1 Agent (with reasoning_effort=high) achieved 43.5% accuracy. </li></ul>However, performance improved noticeably when provided with pre-parsed versions of pages using Databricks&#x27;<a href="https://venturebeat.com/data-infrastructure/databricks-pdf-parsing-for-agentic-ai-is-still-unsolved-new-tool-replaces"> ai_parse_document</a>, indicating that the poor raw PDF performance stems from LLM APIs struggling with parsing rather than reasoning. Even with parsed documents, the experiments show room for improvement.When provided with documents parsed using Databricks&#x27; ai_parse_document:<ul><li>Claude Opus 4.5 Agent achieved 67.8% accuracy (a +30.4 percentage point improvement)</li><li>GPT-5.1 Agent achieved a 52.8% accuracy (a +9.3 percentage point improvement)</li></ul><h2>Three findings that matter for enterprise deployments</h2>The testing identified critical insights for practitioners:Parsing remains the fundamental blocker: Complex tables with nested headers, merged cells and unusual formatting frequently produce misaligned values. Even when given exact oracle pages, agents struggled primarily due to parsing errors, although performance roughly doubled with pre-parsed documents.Document versioning creates ambiguity: Financial and regulatory documents get revised and reissued, meaning multiple valid answers exist depending on the publication date. Agents often stop searching once they find a plausible answer, missing more authoritative sources.Visual reasoning is a gap: About 3% of questions require chart or graph interpretation, where current agents consistently fail. For enterprises where data visualizations communicate critical insights, this represents a meaningful capability limitation.<h2>How enterprises can use OfficeQA</h2>The benchmark&#x27;s design enables specific improvement paths beyond simple scoring. &quot;Since you&#x27;re able to look at the right answer, it&#x27;s easy to tell if the error is coming from parsing,&quot; Elsen explained. This automated evaluation enables rapid iteration on parsing pipelines. The verified ground truth answers also enable RL training similar to coding benchmarks, since there&#x27;s no human judgment required.Elsen said the benchmark provides &quot;a really strong feedback signal&quot; for developers working on search solutions. However, he cautioned against treating it as training data.&quot;At least in my imagination, the goal of releasing this is more as an eval and not as a source of raw training data,&quot; he said. &quot;If you tune too specifically into this environment, then it&#x27;s not clear how generalizable your agent results would be.&quot;<h2>What this means for enterprise AI deployments</h2>For enterprises currently deploying or planning document-heavy AI agent systems, OfficeQA provides a sobering reality check. Even the latest frontier models achieve only 43% accuracy on unprocessed PDFs and fall short of 70% accuracy even with optimal document parsing. Performance on the hardest questions plateaus at 40%, indicating substantial room for improvement.Three immediate implications:Evaluate your document complexity: If your documents resemble the complexity profile of Treasury Bulletins (scanned images, nested table structures, cross-document references), expect accuracy well below vendor marketing claims. Test on your actual documents before production deployment.Plan for the parsing bottleneck: The test results indicate that parsing remains a fundamental blocker. Budget time and resources for custom parsing solutions rather than assuming off-the-shelf OCR will suffice.

Plan for hard question failure modes: Even with optimal parsing, agents plateau at 40% on complex multi-step questions. For mission-critical document workflows that require multi-document analysis, statistical calculations or visual reasoning, current agent capabilities may not be ready without significant human oversight.For enterprises looking to lead in AI-powered document intelligence, this benchmark provides a concrete evaluation framework and identifies specific capability gaps that need solving.

تكشف الأبحاث الأخيرة من Databricks أنه بينما تؤدي وكلاء الذكاء الاصطناعي بشكل جيد في الاختبارات المجردة، فإنهم يواجهون صعوبات كبيرة مع الوثائق المؤسسية، حيث يحققون أقل من 45% من الدقة في المهام التي تعكس أعباء العمل الحقيقية. وهذا يبرز انفصالًا بين المعايير الأكاديمية واحتياجات الأعمال العملية.

La reciente investigación de Databricks revela que, aunque los agentes de IA se desempeñan bien en pruebas abstractas, tienen dificultades significativas con documentos empresariales, logrando menos del 45% de precisión en tareas que reflejan cargas de trabajo del mundo real. Esto destaca una desconexión entre los benchmarks académicos y las necesidades prácticas de las empresas.

La recherche récente de Databricks révèle que, bien que les agents d'IA réussissent bien aux tests abstraits, ils rencontrent des difficultés significatives avec les documents d'entreprise, atteignant moins de 45 % de précision sur des tâches reflétant les charges de travail réelles. Cela met en évidence un décalage entre les benchmarks académiques et les besoins pratiques des entreprises.

Databricks' recent research reveals that while AI agents perform well on abstract tests, they struggle significantly with enterprise documents, achieving less than 45% accuracy on tasks that reflect real-world workloads. This highlights a disconnect between academic benchmarks and the practical needs of businesses.

Databricks' OfficeQA uncovers disconnect: AI agents ace abstract tests but stall at 45% on enterprise docs

Chinese AI startup Zhipu AI aka <a href="https://z.ai/blog/glm-4.6v">Z.ai has released its GLM-4.6V series</a>, a new generation of open-source vision-language models (VLMs) optimized for multimodal reasoning, frontend automation, and high-efficiency deployment. The release includes two models in &quot;large&quot; and &quot;small&quot; sizes: <ol><li>GLM-4.6V (106B), a larger 106-billion parameter model aimed at cloud-scale inference</li><li>GLM-4.6V-Flash (9B), a smaller model of only 9 billion parameters designed for low-latency, local applications</li></ol>Recall that generally speaking, models with more parameters — or internal settings governing their behavior, i.e. weights and biases — are more powerful, performant, and capable of performing at a higher general level across more varied tasks.However, smaller models can offer better efficiency for edge or real-time applications where latency and resource constraints are critical.The defining innovation in this series is the introduction of native function calling in a vision-language model—enabling direct use of tools such as search, cropping, or chart recognition with visual inputs. With a 128,000 token context length (equivalent to a 300-page novel&#x27;s worth of text exchanged in a single input/output interaction with the user) and state-of-the-art (SoTA) results across more than 20 benchmarks, the GLM-4.6V series positions itself as a highly competitive alternative to both closed and open-source VLMs. It&#x27;s available in the following formats:<ul><li><a href="https://docs.z.ai/guides/vlm/glm-4.6v">API access</a> via OpenAI-compatible interface</li><li><a href="https://chat.z.ai">Try the demo</a> on Zhipu’s web interface</li><li><a href="https://huggingface.co/collections/zai-org/glm-46v">Download weights</a> from Hugging Face</li><li>Desktop assistant app available on <a href="https://huggingface.co/spaces/zai-org/GLM-4.5V-Demo-App">Hugging Face Spaces</a></li></ul><h2>Licensing and Enterprise Use</h2>GLM‑4.6V and GLM‑4.6V‑Flash are distributed under the <a href="https://opensource.org/licenses/MIT">MIT license</a>, a permissive open-source license that allows free commercial and non-commercial use, modification, redistribution, and local deployment without obligation to open-source derivative works. This licensing model makes the series suitable for enterprise adoption, including scenarios that require full control over infrastructure, compliance with internal governance, or air-gapped environments.Model weights and documentation are publicly hosted on <a href="https://huggingface.co/collections/zai-org/glm-46v">Hugging Face</a>, with supporting code and tooling available on <a href="https://github.com/zai-org/GLM-V">GitHub</a>. The MIT license ensures maximum flexibility for integration into proprietary systems, including internal tools, production pipelines, and edge deployments.<h2>Architecture and Technical Capabilities</h2>The GLM-4.6V models follow a conventional encoder-decoder architecture with significant adaptations for multimodal input. Both models incorporate a Vision Transformer (ViT) encoder—based on AIMv2-Huge—and an MLP projector to align visual features with a large language model (LLM) decoder. Video inputs benefit from 3D convolutions and temporal compression, while spatial encoding is handled using 2D-RoPE and bicubic interpolation of absolute positional embeddings.A key technical feature is the system’s support for arbitrary image resolutions and aspect ratios, including wide panoramic inputs up to 200:1. In addition to static image and document parsing, GLM-4.6V can ingest temporal sequences of video frames with explicit timestamp tokens, enabling robust temporal reasoning.On the decoding side, the model supports token generation aligned with function-calling protocols, allowing for structured reasoning across text, image, and tool outputs. This is supported by extended tokenizer vocabulary and output formatting templates to ensure consistent API or agent compatibility.<h2>Native Multimodal Tool Use</h2>GLM-4.6V introduces native multimodal function calling, allowing visual assets—such as screenshots, images, and documents—to be passed directly as parameters to tools. This eliminates the need for intermediate text-only conversions, which have historically introduced information loss and complexity.The tool invocation mechanism works bi-directionally:<ul><li>Input tools can be passed images or videos directly (e.g., document pages to crop or analyze).</li><li>Output tools such as chart renderers or web snapshot utilities return visual data, which GLM-4.6V integrates directly into the reasoning chain.</li></ul>In practice, this means GLM-4.6V can complete tasks such as:<ul><li>Generating structured reports from mixed-format documents</li><li>Performing visual audit of candidate images</li><li>Automatically cropping figures from papers during generation</li><li>Conducting visual web search and answering multimodal queries</li></ul><h2>High Performance Benchmarks Compared to Other Similar-Sized Models</h2>GLM-4.6V was evaluated across more than 20 public benchmarks covering general VQA, chart understanding, OCR, STEM reasoning, frontend replication, and multimodal agents. According to the benchmark chart released by Zhipu AI:<ul><li>GLM-4.6V (106B) achieves SoTA or near-SoTA scores among open-source models of comparable size (106B) on MMBench, MathVista, MMLongBench, ChartQAPro, RefCOCO, TreeBench, and more.</li><li>GLM-4.6V-Flash (9B) outperforms other lightweight models (e.g., Qwen3-VL-8B, GLM-4.1V-9B) across almost all categories tested.</li><li>The 106B model’s 128K-token window allows it to outperform larger models like Step-3 (321B) and Qwen3-VL-235B on long-context document tasks, video summarization, and structured multimodal reasoning.</li></ul>Example scores from the leaderboard include:<ul><li>MathVista: 88.2 (GLM-4.6V) vs. 84.6 (GLM-4.5V) vs. 81.4 (Qwen3-VL-8B)</li><li>WebVoyager: 81.0 vs. 68.4 (Qwen3-VL-8B)</li><li>Ref-L4-test: 88.9 vs. 89.5 (GLM-4.5V), but with better grounding fidelity at 87.7 (Flash) vs. 86.8</li></ul>Both models were evaluated using the vLLM inference backend and support SGLang for video-based tasks.<h2>Frontend Automation and Long-Context Workflows</h2>Zhipu AI emphasized GLM-4.6V’s ability to support frontend development workflows. The model can:<ul><li>Replicate pixel-accurate HTML/CSS/JS from UI screenshots</li><li>Accept natural language editing commands to modify layouts</li><li>Identify and manipulate specific UI components visually</li></ul>This capability is integrated into an end-to-end visual programming interface, where the model iterates on layout, design intent, and output code using its native understanding of screen captures.In long-document scenarios, GLM-4.6V can process up to 128,000 tokens—enabling a single inference pass across:<ul><li>150 pages of text (input)</li><li>200 slide decks</li><li>1-hour videos</li></ul>Zhipu AI reported successful use of the model in financial analysis across multi-document corpora and in summarizing full-length sports broadcasts with timestamped event detection.<h2>Training and Reinforcement Learning</h2>The model was trained using multi-stage pre-training followed by supervised fine-tuning (SFT) and reinforcement learning (RL). Key innovations include:<ul><li>Curriculum Sampling (RLCS): Dynamically adjusts the difficulty of training samples based on model progress</li><li>Multi-domain reward systems: Task-specific verifiers for STEM, chart reasoning, GUI agents, video QA, and spatial grounding</li><li>Function-aware training: Uses structured tags (e.g., &lt;think&gt;, &lt;answer&gt;, &lt;|begin_of_box|&gt;) to align reasoning and answer formatting</li></ul>The reinforcement learning pipeline emphasizes verifiable rewards (RLVR) over human feedback (RLHF) for scalability, and avoids KL/entropy losses to stabilize training across multimodal domains<h2>Pricing (API)</h2>Zhipu AI offers competitive pricing for the GLM-4.6V series, with both the flagship model and its lightweight variant positioned for high accessibility.<ul><li>GLM-4.6V: $0.30 (input) / $0.90 (output) per 1M tokens</li><li>GLM-4.6V-Flash: Free</li></ul>Compared to major vision-capable and text-first LLMs, GLM-4.6V is among the most cost-efficient for multimodal reasoning at scale. Below is a comparative snapshot of pricing across providers:USD per 1M tokens — sorted lowest → highest total cost<table><tbody><tr><td>Model</td><td>Input</td><td>Output</td><td>Total Cost</td><td>Source</td></tr><tr><td>Qwen 3 Turbo</td><td>$0.05</td><td>$0.20</td><td>$0.25</td><td><a href="https://www.alibabacloud.com/en/campaign/qwen-ai-landing-page?_p_lc=1&amp;src=qwenai">Alibaba Cloud</a></td></tr><tr><td>ERNIE 4.5 Turbo</td><td>$0.11</td><td>$0.45</td><td>$0.56</td><td><a href="https://cloud.baidu.com/doc/WENXINWORKSHOP/s/Blfmc9do4">Qianfan</a></td></tr><tr><td>GLM‑4.6V</td><td>$0.30</td><td>$0.90</td><td>$1.20</td><td><a href="https://docs.z.ai/guides/overview/pricing">Z.AI</a></td></tr><tr><td>Grok 4.1 Fast (reasoning)</td><td>$0.20</td><td>$0.50</td><td>$0.70</td><td><a href="https://docs.x.ai/docs/models?cluster=us-east-1#detailed-pricing-for-all-grok-models">xAI</a></td></tr><tr><td>Grok 4.1 Fast (non-reasoning)</td><td>$0.20</td><td>$0.50</td><td>$0.70</td><td><a href="https://docs.x.ai/docs/models?cluster=us-east-1#detailed-pricing-for-all-grok-models">xAI</a></td></tr><tr><td>deepseek-chat (V3.2-Exp)</td><td>$0.28</td><td>$0.42</td><td>$0.70</td><td><a href="https://api-docs.deepseek.com/quick_start/pricing">DeepSeek</a></td></tr><tr><td>deepseek-reasoner (V3.2-Exp)</td><td>$0.28</td><td>$0.42</td><td>$0.70</td><td><a href="https://api-docs.deepseek.com/quick_start/pricing">DeepSeek</a></td></tr><tr><td>Qwen 3 Plus</td><td>$0.40</td><td>$1.20</td><td>$1.60</td><td><a href="https://www.alibabacloud.com/en/campaign/qwen-ai-landing-page?_p_lc=1&amp;src=qwenai">Alibaba Cloud</a></td></tr><tr><td>ERNIE 5.0</td><td>$0.85</td><td>$3.40</td><td>$4.25</td><td><a href="https://cloud.baidu.com/doc/WENXINWORKSHOP/s/Blfmc9do4">Qianfan</a></td></tr><tr><td>Qwen-Max</td><td>$1.60</td><td>$6.40</td><td>$8.00</td><td><a href="https://www.alibabacloud.com/en/campaign/qwen-ai-landing-page?_p_lc=1&amp;src=qwenai">Alibaba Cloud</a></td></tr><tr><td>GPT-5.1</td><td>$1.25</td><td>$10.00</td><td>$11.25</td><td><a href="https://openai.com/pricing">OpenAI</a></td></tr><tr><td>Gemini 2.5 Pro (≤200K)</td><td>$1.25</td><td>$10.00</td><td>$11.25</td><td><a href="https://ai.google.dev/gemini-api/docs/pricing">Google</a></td></tr><tr><td>Gemini 3 Pro (≤200K)</td><td>$2.00</td><td>$12.00</td><td>$14.00</td><td><a href="https://ai.google.dev/gemini-api/docs/pricing">Google</a></td></tr><tr><td>Gemini 2.5 Pro (&gt;200K)</td><td>$2.50</td><td>$15.00</td><td>$17.50</td><td><a href="https://ai.google.dev/gemini-api/docs/pricing">Google</a></td></tr><tr><td>Grok 4 (0709)</td><td>$3.00</td><td>$15.00</td><td>$18.00</td><td><a href="https://docs.x.ai/docs/models?cluster=us-east-1#detailed-pricing-for-all-grok-models">xAI</a></td></tr><tr><td>Gemini 3 Pro (&gt;200K)</td><td>$4.00</td><td>$18.00</td><td>$22.00</td><td><a href="https://ai.google.dev/gemini-api/docs/pricing">Google</a></td></tr><tr><td>Claude Opus 4.1</td><td>$15.00</td><td>$75.00</td><td>$90.00</td><td><a href="https://docs.anthropic.com/claude/docs/models-overview">Anthropic</a></td></tr></tbody></table><h2>Previous Releases: GLM‑4.5 Series and Enterprise Applications</h2>Prior to GLM‑4.6V, Z.ai released the GLM‑4.5 family in mid-2025, establishing the company as a serious contender in open-source LLM development. The flagship GLM‑4.5 and its smaller sibling GLM‑4.5‑Air both support reasoning, tool use, coding, and agentic behaviors, while offering strong performance across standard benchmarks. The models introduced dual reasoning modes (“thinking” and “non-thinking”) and could automatically generate complete PowerPoint presentations from a single prompt — a feature positioned for use in enterprise reporting, education, and internal comms workflows. Z.ai also extended the GLM‑4.5 series with additional variants such as GLM‑4.5‑X, AirX, and Flash, targeting ultra-fast inference and low-cost scenarios.Together, these features position the GLM‑4.5 series as a cost-effective, open, and production-ready alternative for enterprises needing autonomy over model deployment, lifecycle management, and integration pipel<h2>Ecosystem Implications</h2>The GLM-4.6V release represents a notable advance in open-source multimodal AI. While large vision-language models have proliferated over the past year, few offer:<ul><li>Integrated visual tool usage</li><li>Structured multimodal generation</li><li>Agent-oriented memory and decision logic</li></ul>Zhipu AI’s emphasis on “closing the loop” from perception to action via native function calling marks a step toward agentic multimodal systems. The model’s architecture and training pipeline show a continued evolution of the GLM family, positioning it competitively alongside offerings like OpenAI’s GPT-4V and Google DeepMind’s Gemini-VL.<h2>Takeaway for Enterprise Leaders</h2>With GLM-4.6V, Zhipu AI introduces an open-source VLM capable of native visual tool use, long-context reasoning, and frontend automation. It sets new performance marks among models of similar size and provides a scalable platform for building agentic, multimodal AI systems.

أطلقت شركة Zhipu AI الصينية، المعروفة أيضًا باسم Z.ai، سلسلة GLM-4.6V، وهي جيل جديد من نماذج اللغة-الرؤية مفتوحة المصدر، المحسّنة للتفكير متعدد الوسائط والنشر الفعال. تشمل السلسلة نموذجين: GLM-4.6V الذي يحتوي على 106 مليار معلمة للتفسير على نطاق السحابة وGLM-4.6V-Flash الذي يحتوي على 9 مليارات معلمة للتطبيقات ذات الكمون المنخفض.

La startup china Zhipu AI, conocida también como Z.ai, ha lanzado su serie GLM-4.6V, una nueva generación de modelos de lenguaje-visual de código abierto optimizados para el razonamiento multimodal y el despliegue eficiente. La serie incluye dos modelos: GLM-4.6V con 106 mil millones de parámetros para inferencia a escala en la nube y GLM-4.6V-Flash con 9 mil millones de parámetros para aplicaciones de baja latencia.

La startup chinoise Zhipu AI, également connue sous le nom de Z.ai, a lancé sa série GLM-4.6V, une nouvelle génération de modèles de vision-langage open-source optimisés pour le raisonnement multimodal et le déploiement efficace. La série comprend deux modèles : GLM-4.6V avec 106 milliards de paramètres pour l'inférence à l'échelle du cloud et GLM-4.6V-Flash avec 9 milliards de paramètres pour des applications à faible latence.

Chinese AI startup Zhipu AI, also known as Z.ai, has launched its GLM-4.6V series, a new generation of open-source vision-language models optimized for multimodal reasoning and efficient deployment. The series includes two models: GLM-4.6V with 106 billion parameters for cloud-scale inference and GLM-4.6V-Flash with 9 billion parameters for low-latency applications.

Z.ai debuts open source GLM-4.6V, a native tool-calling vision model for multimodal reasoning

<a href="https://anthropic.com/">Anthropic</a> has launched a beta integration that connects its fast-growing <a href="https://www.claude.com/product/claude-code">Claude Code</a> programming agent <a href="https://www.anthropic.com/news/claude-code-and-slack">directly into Slack</a>, allowing software engineers to delegate coding tasks without leaving the workplace messaging platform where much of their daily communication already happens.The release, which Anthropic describes as a &quot;<a href="https://www.anthropic.com/news/claude-code-and-slack">research preview</a>,&quot; is the company&#x27;s latest move to embed its technology deeper into enterprise workflows — and comes as Claude Code has emerged as a surprise revenue engine, generating more than $1 billion in annualized revenue just six months after its public debut.&quot;The critical context around engineering work often lives in Slack, including bug reports, feature requests and engineering discussions,&quot; the company wrote in a <a href="https://www.anthropic.com/news/claude-code-and-slack">blog post</a>. &quot;When a bug report appears or a teammate needs a code fix, you can now tag Claude in Slack to automatically spin up a Claude Code session using the surrounding context.&quot;<h2>From bug report to pull request: How the new Slack integration actually works</h2>The mechanics are deceptively simple but address a persistent friction point in software development: The gap between where problems are discussed and where they are fixed.When a user mentions @Claude in a Slack channel or thread, Claude analyzes the message to determine whether it constitutes a coding task. If it does, the system automatically creates a new <a href="https://www.claude.com/product/claude-code">Claude Code</a> session. Users can also explicitly instruct Claude to treat requests as coding tasks.Claude gathers context from recent Slack channel and thread messages to feed into the Claude Code session. It will use this context to automatically choose which repository to run the task on based on the repositories that have been authenticated to Claude Code on the web.As the Claude Code session progresses, Claude posts status updates back to the Slack thread. Once complete, users receive a link to the full session where they can review changes, along with a direct link to open a pull request.The feature builds on Anthropic&#x27;s existing <a href="https://www.claude.com/claude-and-slack">Claude for Slack</a> integration and requires users to have access to Claude Code on the web. In practical terms, a product manager reporting a bug in Slack could tag Claude, which would then analyze the conversation context, identify the relevant code repository, investigate the issue, propose a fix and post a pull request — all while updating the original Slack thread with its progress.<h2>Why Anthropic is betting big on enterprise workflow integrations</h2>The Slack integration arrives at a pivotal moment for Anthropic. Claude Code has already hit <a href="https://www.linkedin.com/posts/mikekrieger_anthropic-acquires-bun-as-claude-code-reaches-activity-7401686833233076224-n3TI/">$1 billion in revenue</a>, six months after its public debut, according to a LinkedIn post from Anthropic&#x27;s CPO Mike Krieger. The coding agent continues to barrel toward scale with customers like Netflix, Spotify and Salesforce.The velocity of that growth helps explain why Anthropic made its <a href="https://bun.com/blog/bun-joins-anthropic">first-ever acquisition</a> earlier this month, of developer tool startup Bun. Anthropic declined to comment on specific financial details. <a href="https://bun.com/blog/bun-joins-anthropic">Bun</a> is a breakthrough JavaScript runtime that claims to be dramatically faster than the leading competition. As an all-in-one toolkit — combining runtime, package manager, bundler and test runner — it&#x27;s become essential infrastructure for AI-led software engineering, helping developers build and test applications at unprecedented velocity.Since becoming generally available in May 2025, <a href="https://www.claude.com/product/claude-code">Claude Code</a> has grown from its origins as an internal engineering experiment into a critical tool for many of the world&#x27;s category-leading enterprises, including <a href="https://www.netflix.com/">Netflix</a>, <a href="https://open.spotify.com/">Spotify</a>, <a href="https://kpmg.com/us/en.html">KPMG</a>, <a href="https://www.loreal.com/en/">L&#x27;Oreal</a> and <a href="https://www.salesforce.com/">Salesforce</a> — and <a href="https://bun.com/blog/bun-joins-anthropic">Bun</a> has been key in helping scale its infrastructure throughout that evolution.The acquisition signals that Anthropic views <a href="https://www.claude.com/product/claude-code">Claude Code</a> not as a peripheral feature but as a core business line worth substantial investment. The Slack integration extends that bet, positioning Claude Code as an ambient presence in the workspaces where engineering decisions actually get made.According to an Anthropic spokesperson, companies including <a href="https://www.rakuten.com/">Rakuten</a>, <a href="https://www.novonordisk.com/">Novo Nordisk</a>, <a href="https://www.uber.com/">Uber</a>, <a href="https://www.snowflake.com/en/">Snowflake</a> and <a href="https://ramp.com/business-cards?utm_campaign_id=15309675834&amp;utm_ad_group_id=134786044053&amp;utm_ad_id=737002215776&amp;utm_matchtype=e&amp;utm_term=ramp&amp;utm_campaign=google-search-branded&amp;utm_source=google&amp;utm_medium=cpc&amp;hsa_acc=5740001625&amp;hsa_cam=15309675834&amp;hsa_grp=134786044053&amp;hsa_ad=737002215776&amp;hsa_src=g&amp;hsa_tgt=aud-2197767270940:kwd-27557331&amp;hsa_kw=ramp&amp;hsa_mt=e&amp;hsa_net=adwords&amp;hsa_ver=3&amp;cq_cmp=15309675834&amp;cq_con=134786044053&amp;cq_plac=&amp;cq_net=g&amp;cq_plt=gp&amp;gad_source=1&amp;gad_campaignid=15309675834&amp;gbraid=0AAAAACjAjKPV4x9hvhWDm_aa7C0JiphLP&amp;gclid=Cj0KCQiAi9rJBhCYARIsALyPDts_kz5iHuPDKNlcI4ZU2sHwiscemYdv3EbJXX7uKXC2wnh2yc5C1QAaAi2WEALw_wcB">Ramp</a> now use Claude Code for both professional and novice developers. Rakuten, the Japanese e-commerce giant, has reportedly reduced software development timelines from 24 to 5 days using the tool — a 79% reduction that illustrates the productivity claims Anthropic has been making.<h2>Claude Code&#x27;s rapid rise from internal experiment to billion-dollar product</h2>The Slack launch is the latest in a rapid series of Claude Code expansions. In late November, <a href="https://www.anthropic.com/news/claude-opus-4-5">Claude Code was added to Anthropic&#x27;s desktop apps</a>, including the Mac version. Previously, Claude Code was limited to mobile apps and the web. The desktop version allows software engineers to code, research and update work with multiple local and remote sessions running at the same time.That release accompanied Anthropic&#x27;s unveiling of Claude <a href="https://www.anthropic.com/news/claude-opus-4-5">Opus 4.5</a>, its newest and most capable model. Claude Opus 4.5 is available today on the company&#x27;s apps, API and on all three major cloud platforms. Pricing is $5/$25 per million tokens — making Opus-level capabilities accessible to even more users, teams and enterprises.The company has also invested heavily in the developer infrastructure that powers Claude Code. In late November, <a href="https://www.anthropic.com/engineering/advanced-tool-use">Anthropic released three new beta features for tool use</a>: Tool Search Tool, which allows Claude to access thousands of tools without consuming its context window; Programmatic Tool Calling, which allows Claude to invoke tools in a code execution environment, thus reducing the impact on the model&#x27;s context window; and Tool Use Examples, which provides a universal standard for demonstrating how to effectively use a given tool.The <a href="https://modelcontextprotocol.io/docs/getting-started/intro">Model Context Protocol (MCP)</a> is an open standard for connecting AI agents to external systems. Connecting agents to tools and data traditionally requires a custom integration for each pairing, creating fragmentation and duplicated effort that makes it difficult to scale truly connected systems. MCP provides a universal protocol — developers implement MCP once in their agent, and it unlocks an entire ecosystem of integrations.<h2>Inside Anthropic&#x27;s own AI transformation: What happens when engineers use Claude all day</h2>Anthropic has been unusually transparent about how its own engineers use Claude Code — and the findings offer a preview of broader workforce implications. In August 2025, Anthropic<a href="https://www.anthropic.com/research/how-ai-is-transforming-work-at-anthropic"> surveyed 132 engineers and researchers</a>, conducted 53 in-depth qualitative interviews and studied internal Claude Code usage data to understand how AI use is changing work at the company.Employees self-reported <a href="https://www.anthropic.com/research/how-ai-is-transforming-work-at-anthropic">using Claude in 60% of their work</a> and achieved a 50% productivity boost, a 2-3x increase from this time last year. This productivity looks like slightly less time per task category, but considerably more output volume.Perhaps most notably, 27% of Claude-assisted work consists of tasks that wouldn&#x27;t have been done otherwise, such as scaling projects, making nice-to-have tools like interactive data dashboards, and exploratory work that wouldn&#x27;t be cost-effective if done manually.The internal research also revealed how Claude is changing the nature of engineering collaboration. The maximum number of consecutive tool calls Claude Code makes per transcript increased by 116%. Claude now chains together 21.2 independent tool calls without the need for human intervention, versus 9.8 tool calls from six months ago.The number of human turns decreased by 33%. The average number of human turns decreased from 6.2 to 4.1 per transcript, suggesting that less human input is necessary to accomplish a given task now, compared to six months ago.But the research also surfaced tensions. One prominent theme was that Claude has become the first stop for questions that once went to colleagues. &quot;It has reduced my dependence on [my team] by 80%, [but] the last 20% is crucial, and I go and talk to them,&quot; one engineer explained. Several engineers said they &quot;bounce ideas off&quot; Claude, similar to interactions with human collaborators.Some appreciate the reduced social friction, but others resist the change or miss the older way of working: &quot;I like working with people, and it is sad that I need them less now.&quot;<h2>How Anthropic stacks up against OpenAI, Google and Microsoft in the enterprise AI race</h2>Anthropic is not alone in racing to capture the enterprise coding market. <a href="https://openai.com/">OpenAI</a>, <a href="https://www.google.com/">Google</a> and <a href="https://www.microsoft.com/en-us/">Microsoft</a> (through <a href="https://github.com/features/copilot">GitHub Copilot</a>) are all pursuing similar integrations. The Slack launch gives Anthropic a presence in one of the most widely used enterprise communication platforms — Slack claims over 750,000 organizations use its software.The deal comes as Anthropic pursues a more disciplined growth path than rival OpenAI, focusing on enterprise customers and coding workloads. Internal financials reported by The Wall Street Journal show that <a href="https://www.wsj.com/tech/ai/openai-anthropic-profitability-e9f5bcd6?gaa_at=eafs&amp;gaa_n=AWEtsqf_yQrKzPR_c7Gjwi4FvVkM-Uc0I6k_V6Kev-0xg4RiUud_8rUKLIBVacKJN5E%3D&amp;gaa_ts=693721bf&amp;gaa_sig=m9SvmJAAB8gRVI39wbvEYOLFa9vd_lQbtYoQE1ROCKxaf6CXTOxXgKn4HmpIwymS0BUt58auzZusgCjq74SM2w%3D%3D">Anthropic expects to break even by 2028</a> — two years earlier than OpenAI, which continues to invest heavily in infrastructure as it expands into video, hardware and consumer products.The move also marks an increased push into developer tooling. Anthropic has recently seen backing from some of tech&#x27;s biggest titans. Microsoft and Nvidia pledged up to <a href="https://www.cnbc.com/2025/12/03/anthropic-claude-reportedly-preparing-ipo-race-openai-chatgpt-ft-wilson-sonsini-goodrich-rosati.html">$15 billion in fresh investment</a> in Anthropic last month, alongside a $30 billion commitment from Anthropic to run Claude Code on Microsoft&#x27;s cloud. This is in addition to the <a href="https://www.anthropic.com/news/anthropic-amazon-trainium">$8 billion invested by Amazon </a>and <a href="https://www.cnbc.com/2025/12/04/google-replit-ai-vibe-coding-anthropic-cursor.html">$3 billion by Google</a>.The cross-investment from both <a href="https://microsoft.com/">Microsoft</a> and <a href="https://www.google.com/">Google</a> — fierce competitors in the cloud and AI spaces — highlights Anthropic&#x27;s valuable enterprise positioning. By integrating with Slack (which is owned by Salesforce), Anthropic further embeds itself in the enterprise software ecosystem while remaining platform-agnostic.<h2>What the Slack integration means for developers — and whether they can trust it</h2>For engineering teams, the Slack integration promises to collapse the distance between problem identification and resolution. A bug report in a Slack channel can immediately trigger an investigation. A feature request can spawn a prototype. A code review comment can generate a refactor.But the integration also raises questions about oversight and code quality. Most Anthropic employees use Claude frequently while reporting they can &quot;fully delegate&quot; only 0 to 20% of their work to it. Claude is a constant collaborator, but using it generally involves active supervision and validation, especially in high-stakes work, versus handing off tasks requiring no verification at all.Some employees are concerned about the atrophy of deeper skillsets required for both writing and critiquing code — &quot;When producing output is so easy and fast, it gets harder and harder to actually take the time to learn something.&quot;The Slack integration, by making Claude Code invocation as simple as an @mention, may accelerate both the productivity benefits and the skill-atrophy concerns that Anthropic&#x27;s own research has documented.<h2>The future of coding may be conversational—and Anthropic is racing to prove it</h2>The beta launch marks the beginning of what Anthropic expects will be a broader rollout, with documentation forthcoming for teams looking to deploy the integration and refinements planned based on user feedback during the research preview phase.For Anthropic, the Slack integration is a calculated bet on a fundamental shift in how software gets written. The company is wagering that the future of coding will be conversational — the walls between where developers talk about problems and where they solve them will dissolve entirely. The companies that win enterprise AI, in this view, will be the ones that meet developers not in specialized tools but in the chat windows they already have open all day.Whether that vision becomes reality will depend on whether <a href="https://www.claude.com/product/claude-code">Claude Code</a> can deliver enterprise-grade reliability while maintaining the security that organizations demand. The early returns are promising: A billion dollars in revenue, a roster of Fortune 500 customers and a growing ecosystem of integrations suggest Anthropic is onto something real.But in one of Anthropic&#x27;s own internal interviews, an engineer offered a more cautious assessment of the transformation underway: &quot;Nobody knows what&#x27;s going to happen… the important thing is to just be really adaptable.&quot;In the age of AI coding agents, that may be the only career advice that holds up.

أطلقت شركة أنثروبيك تكاملًا تجريبيًا لوكيل البرمجة كلود كود مع سلاك، مما يتيح لمهندسي البرمجيات إدارة مهام البرمجة مباشرة داخل منصة المراسلة. يهدف هذا التكامل إلى تبسيط سير العمل من خلال السماح للمهندسين بالوصول إلى سياقات حاسمة، مثل تقارير الأخطاء وطلبات الميزات، دون الحاجة إلى تغيير التطبيقات.

Anthropic ha lanzado una integración beta de su agente de programación Claude Code con Slack, permitiendo a los ingenieros de software gestionar tareas de codificación directamente dentro de la plataforma de mensajería. Esta integración busca optimizar los flujos de trabajo al permitir a los ingenieros acceder a contextos críticos, como informes de errores y solicitudes de funciones, sin cambiar de aplicación.

Anthropic a lancé une intégration bêta de son agent de programmation Claude Code avec Slack, permettant aux ingénieurs logiciels de gérer des tâches de codage directement au sein de la plateforme de messagerie. Cette intégration vise à rationaliser les flux de travail en permettant aux ingénieurs d'accéder à des contextes critiques, tels que les rapports de bogues et les demandes de fonctionnalités, sans changer d'application.

Anthropic has introduced a beta integration of its Claude Code programming agent with Slack, enabling software engineers to manage coding tasks directly within the messaging platform. This integration aims to streamline workflows by allowing engineers to access critical context, such as bug reports and feature requests, without switching applications.