Anthropic

IBM has been living for decades that no one could kill COBOL. Anthropic has other plans

by

IBM shares fell about 13.2% yesterday on the New York Stock Exchange for a simple reason: Anthropic advertisement that its AI model, Claude, can be used to modernize systems that are based on the legendary COBOL programming language. And that is something that seemed virtually impossible. The immortal language. As Anthropic itself indicates, it is estimated that COBOL manages 95% of all transactions made at ATMs in the US. A 2022 study revealed that there are 800 billion lines of COBOL code that continue to operate in production systems on a daily basis. That almost no one uses anymore. Faced with this reality is another equally powerful one: almost no one programs in COBOL anymore, because this language has been with us for 65 years and has ended up being replaced by modern programming languages. The question, of course, is who is in charge of those millions of lines of code if there are almost no human programmers who can do it. Anthropic itself made it clear: “the number of people who understand COBOL decreases every year.” AI to the rescue. That’s where Claude, Anthropic’s family of generative AI models, comes in. According to this company, Claude is now capable of “modernizing” COBOL despite how difficult and expensive it was to carry out something like that. IBM has been trying for years and in fact applied that same recipebut its AI (Watson) does not seem to have managed too much progress. Claude helps, but there must be a human expert supervising. At Anthropic they promise that their AI model is capable of reading the entire code base of a COBOL project, identifying entry points, execution paths through subroutines, mapping data flows and documenting dependencies. They highlight, however, that with the supervision of a human expert this can help modernize and polish all types of COBOL-based systems. Critical systems. Of course, the question is whether AI will actually deliver on that promise, especially when we’re talking about absolutely critical systems used in financial transactions. According to Anthropic “the modernization of the code legacy It has been stagnant for years because understanding it cost more than rewriting it. “AI reverses that equation.” COBOL is no longer IBM’s ace in the hole. It’s hard to know how much of IBM’s business depended on COBOL systems, but it’s certainly a relevant part. In 2025 the company achieved revenue of $67.5 billion. About 45% comes from software. The rest is consulting and infrastructure, and this last division is where the IT business is included. IBM Z mainframesclosely linked to COBOL systems. It’s reasonable to think that revenues dependent on mainframes and COBOl are around 20% of IBM’s revenues (and probably more in profits). AI and the SaaSpocalypse. What happened with IBM and COBOL is the latest case of a software that seemed to have a long-term future but with AI may not have such a long-term future. Investors now seem to think that AI will replace many of these systems and SaaS platforms. It is indeed what has been called “SaaSpocalypse” in reference to the stock market falls of this type of companies in recent months: Salesforce, SAP, Microsoft, Adobe, Intuit and Atlassian have suffered notable falls in the stock market that are around 30-40% on average. But. This investor panic that is being experienced contrasts with the current reality: AI models are proving to be able to do surprising things in the field of programming, but they are far from being perfect. The code must be reviewed, and IBM itself he already made it clear In a 1979 training manual: “A computer can never be held responsible. Therefore, it should never make an administrative decision.” IBM has already survived other crises. The blue giant has suffered a blow to the stock market, but it is one of those technology companies that have managed to recover and resist all the attacks of an industry that is normally merciless. IBM itself also has its modernization solutions for its clients, and some analysts they are clear that in fact IBM will make more money than before if COBOL finally goes away. In Xataka | Old programmers never die, and Silicon Valley is realizing that

Anthropic just accused DeepSeek and other Chinese companies of “distilling” Claude

by

For months we have talked about the race between the United States and China to dominate artificial intelligence as if it were only a question of who trains the most powerful model or launches the next version first. But the pulse begins to move to another, more delicate area: that of the rules of the game. When one laboratory accuses another of extracting capabilities from its system to accelerate its own development, the discussion goes beyond the technical. That’s exactly what Anthropic just did by denounce “distillation” campaigns against his model Claude. The complaint. In a text published this Monday, the company claims to have detected “industrial-scale campaigns” aimed at extracting Claude’s capabilities. According to their version, the activities attributed to DeepSeekMoonshot and MiniMax reportedly involved more than 16 million queries, question and answer interactions, and were channeled through approximately 24,000 fraudulent accounts, in violation of their terms of service and regional access restrictions. The race and the suspicion. The announcement by the firm led by Darío Amodei occurs in a context of growing tension around the progress of Chinese AI. Let us remember that DeepSeek altered the Silicon Valley landscape a year ago with the launch of R1, a competitive model that was presented as Developed at a fraction of the cost of American alternatives. The impact was immediate on the markets and revived the political debate in Washington about the technological advantage over China. Distilling is not always cheating. Anthropic itself recognizes that distillation is a common technique in the sector. It consists, in simple terms, of training a less capable model using the responses generated by a more powerful one, something that large laboratories use to create smaller, cheaper versions of their own systems. The problem, according to the company, appears when this practice is used to “acquire powerful capabilities from other laboratories in a fraction of the time and at a fraction of the cost” that developing them independently would entail. In that case, distillation would cease to be an internal optimization and would become, always according to Anthropic, a way of taking advantage of the work of others. Recognizable pattern. The three laboratories would have used fraudulent accounts and proxy services to access Claude on a large scale while trying to avoid detection systems. The company details infrastructures, what it calls “hydra cluster”, extensive networks of accounts that distribute traffic between its API and third-party cloud platforms, so that when one account was blocked, another took its place. Anthropic maintains that what differentiated these activities from normal use was not an isolated query, but rather the massive and coordinated repetition of requests aimed at extracting very specific capabilities from the model. Three campaigns. Although Anthropic presents the campaigns as part of the same dynamic, it distinguishes relevant nuances. DeepSeek would have focused its more than 150,000 queries on extracting reasoning capabilities and generating safe alternatives to politically sensitive questions. Moonshot, with more than 3.4 million queries, would have been oriented towards the development of agents capable of using tools and manipulating computing environments. MiniMax would concentrate the largest volume, more than 13 million queries, and according to Anthropic’s account, it reacted in a matter of hours to the launch of a new system, redirecting its traffic to try to extract capabilities from its most recent system. A geopolitical issue. The company states that illicitly distilled models may lose safeguards that seek to prevent state or non-state actors from using AI for purposes such as the development of biological weapons or disinformation campaigns. It also argues that distillation undermines export controls by allowing foreign laboratories to close the gap in other ways, while at the same time recognizing that executing these large-scale extractions requires access to advanced chips, thus reinforcing the logic of restricting their availability while, at the same time, remembering that the risk would grow if these capabilities end up being integrated into military, intelligence or surveillance systems. Images | Xataka with Nano Banana Pro In Xataka | Seedance is the greatest brutality we have seen generating video. And it has an uncomfortable message: it has surpassed Sora and Veo without NVIDIA chips

Anthropic corners Gemini 3 Pro and GPT-5.2 more than ever

by

Think for a moment about the artificial intelligence models you have used in recent days. It may have been through ChatGPT, Gemini either Claudeor perhaps through tools like Codex, Claude Code or AI Cursor. In practice, the choice is usually simple: we end up using what best fits what we need at any given moment, almost without stopping to think about the technology behind it. However, that balance changes frequently. Each new model that appears promises improvements, new capabilities or different ways of working, and with it a fairly direct question returns: if it is worth trying, if it can really offer us something better or if what we already use is still enough. Claude Sonnet 4.6 just came to the foreand this is how it is positioned against the competition. Claude Sonnet’s starting point 4.6. Here we find what Anthropic describes as a transversal improvement in capabilities, which includes advances in coding, computer use, long-context reasoning, agent planning, and tasks typical of intellectual and creative work. Added to this set is a context window of up to one million tokens in beta, designed to process entire code bases, extensive contracts or large collections of information without fragmentation. Three levels, the same map. To understand where Sonnet 4.6 fits in, it’s worth looking at how Anthropic tends to organize its family of models into different levels with different objectives. Haiku prioritizes speed and efficiency, Opus is reserved for tasks that require the deepest reasoning, and Sonnet occupies the middle ground, designed as a balance between capacity and operating cost. Within this framework, the company maintains that the new Sonnet comes close in some real jobs to the performance previously associated with the Opus, an ambitious claim. When AI starts using the computer. One of the improvements that Anthropic highlights most strongly in Sonnet 4.6 is its progress in what it calls computer usethat is, the ability of the model to interact with software in a way similar to a person, without depending on APIs designed specifically for automation. This progress is supported by references such as OSWorld-Verified, a testing environment with real applications where the Sonnet family has been improving steadily over several months. The company also recognizes limits and risks that we have talked about before, such as attempts at manipulation through prompt injection. Searching for the ‘best’ model. At this point, the relevant question stops being how much Sonnet 4.6 has improved in absolute terms and begins to focus on how it is compared to the other large models that today compete for the same space of use. The comparison is not simple nor does it allow for a single winner, because each system excels in different areas and responds to different technical priorities. That is why it is advisable to read the benchmarks with a practical perspective, identifying in which specific tasks the real differences appear. Where each model stands out. The direct comparison with GPT-5.2 draws a distribution of strengths rather than a clear victory. According to the table published by Anthropic, Sonnet 4.6 stands out especially widely in the autonomous use of the computer measured in OSWorld-Verified, in addition to showing an advantage in office tasks (GDPval-AA Elo) and in some analysis or problem solving scenarios (Finance Agent v1.1, ARC-AGI-2). GPT-5.2, for its part, maintains better results in graduate-level reasoning (GPQA Diamond), visual comprehension (MMMU-Pro) and terminal programming (Terminal-Bench 2.0), with nuances such as results marked as Pro in some tests (BrowseComp, HLE) or self-reported grades in Terminal-Bench 2.0. The comparison with Gemini 3 Pro introduces a different nuance, because here the advantages are concentrated above all in the field of reasoning and general knowledge. The Google model obtains better results in graduate-level reasoning tests (GPQA Diamond) and in wide-ranging multilingual questionnaires (MMMLU), in addition to being ahead in visual reasoning without tools (MMMU-Pro). Sonnet 4.6, on the other hand, retains a certain advantage when external tools or scenarios closer to the applied work come into play. The absence of some comparable data in the table itself forces, in any case, to interpret this duel with caution. Where Sonnet 4.6 can be used. The new model is available in all Claude plans, including the free level, where it also becomes the default option within claude.ai and Claude Cowork. It can also be used through Claude Code, the API and the main cloud platforms, maintaining the same price as the Sonnet 4.5 version. After going through capabilities, limits and comparisons, the real decision returns to the user’s daily life. Sonnet 4.6 aims to be especially useful in productive tasks, direct interaction with software and long workflows, while GPT-5.2 and Gemini 3 Pro maintain advantages in academic reasoning, visual comprehension or general knowledge depending on the test considered. No one dominates all fronts, and that fragmentation defines the current moment of artificial intelligence. Images | Anthropic In Xataka | In 2025, AI seemed to have hit a wall of progress. A volatilized wall in February 2026 In Xataka | The great revolution of GPT-5.3 Codex and Claude Opus 4.6 is not that they are smarter. It’s that they can improve themselves

Anthropic wanted to secretly scan and then destroy millions of books to train its AI. It hasn’t been so secret

by

A language model for AI needs input if it is to be trained to be more accurate and effective. The issue is how the information is obtained and whether there is an ethical way to do it that is profitable for the technology company in power. There is no doubt that the preferred option for companies has been to use all possible physical and digital content without anyone’s permission. There is also evidence. A judicial leak reveals that Anthropic invested tens of millions of dollars in acquiring and digitizing literary works without permission from the authors. According to account Washington Post, the project, internally called “Panama”, was part of a frenetic race among big technology companies to accumulate massive data to train their artificial intelligence models. How it all started. The Panama Project was launched by Anthropic in early 2024. According to internal documents revealed per the Washington Post, the goal was to “destructively scan every book in the world.” Furthermore, these documents also explicitly state that the company did not want anyone to know that they were working on it. In about a year, the company spent tens of millions of dollars buying millions of books, cutting their spines with hydraulic machines and scanning their pages to feed the AI ​​models that power Claudeits star chatbot. According to the media, the books, once digitized, ended up being recycled. Because has come to light. The details of the project have been revealed in a lawsuit for infringement of rights copyright filed by literary authors against Anthropic. Although the company agreed to pay $1.5 billion to close the case in August 2025, a district judge decided to make more than 4,000 pages of internal documents public last week, exposing the entire operation. They are not the only ones. Court documents reveal that other technology companies such as Meta, Google and OpenAI had also participated in this race to obtain massive information to train their models. According to revealed According to the documents, an Anthropic co-founder theorized in January 2023 that training AI models with books could teach them “how to write well” instead of imitating “low-quality internet slang.” On the other hand, an internal Meta email from 2024 described access to a digital library of books as “essential” to be competitive with rivals in the race to dominate AI. However, the documents revealed by the media also show how Meta employees expressed concern on several occasions about the legality of downloading millions of books without permission. An internal email from December 2023 indicates that the practice had been approved after being “escalated to MZ,” apparently referring to CEO Mark Zuckerberg. According to court records to which the media has had access, the companies did not consider it “practical” to obtain direct permission from publishers and authors. Instead, they found ways to mass-acquire books without the writers’ knowledge, including downloading unauthorized copies from third-party sites. Chat logs from April 2024 show an employee asking why they were using servers rented from Amazon to download torrents instead of Facebook’s own. The answer: “Avoid the risk of tracing” the activity back to the company. Data torrent. The documents to which the Washington Post has had access also they test that Ben Mann, co-founder of Anthropic, personally downloaded over 11 days in June 2021 a collection of books from LibGen, a gigantic library of copyrighted content. The outlet further revealed that, a year later, in July 2022, Mann celebrated the launch of the ‘Pirate Library Mirror’ website, which boasts a massive database of books and openly claims to violate copyright laws. “Just in time!!!” Mann wrote to other Anthropic employees, according to the outlet. Anthropic stated in legal documents that it never trained a revenue-generating business model using LibGen data nor did it use Pirate Library Mirror to train any full model. Anthropic’s legal solution. According to point the medium in its article, faced with the legal risk, Anthropic changed its strategy. The company hired Tom Turvey, a Silicon Valley veteran who had helped create the project Google Books two decades earlier. Under his direction, Anthropic considered purchasing books from libraries or secondhand bookstores, including New York’s iconic Strand bookstore. The company ultimately ended up buying millions of books and stacking them in a giant warehouse, often in batches of tens of thousands, according to court filings. The Washington Post assures In addition, the company worked with used book sellers in the United Kingdom. A project proposal mentions that Anthropic sought to “convert between 500,000 and two million books in a six-month period.” What the law says. Most legal cases against AI companies are still ongoing, but the media mention two court rulings that have considered that the use of books to train AI models without permission from the author or publisher may be legal under the “fair use” doctrine of copyright. In June 2025, District Judge William Alsup determined that Anthropic had the right to use books to train AI models because they process them in a “transformative” way. He compared the process to teachers “teaching schoolchildren to write well.” That same month, Judge Vince Chhabria ruled in the Meta case that the authors had not shown that the company’s AI models could harm the sales of their books. In the Anthropic case, the physical book scanning project was considered legal, but the judge determined that the company may have infringed copyright by downloading millions of books without authorization before launching Project Panama. The final agreement. Instead of facing a trial, Anthropic agreed to pay $1.5 billion to publishers and authors without admitting guilt. According to point According to the media, authors whose books were downloaded can claim their share of the settlement, estimated at about $3,000 per title. Cover image | Emil Widlund and Anthropic In Xataka | If AI is going to leave us without jobs, in the United Kingdom they are already seriously discussing the solution: a universal basic income

Programming is the new board of AI. OpenAI and Anthropic have made it clear with GPT-5.3-Codex and Claude Opus 4.6

by

When ChatGPT broke out in November 2022, OpenAI seemed unrivaled. And, to a large extent, that was the case. That chatbot, despite its errors and limitations, inaugurated a category of its own. However, in the technology sector advantages are rarely permanent and, in 2026, the position of the company led by Sam Altman It’s a far cry from what it had then. Google has managed to attract the general public with Nano Banana Prowhile Gemini steadily gaining ground as an artificial intelligence chatbot. At the same time, ChatGPT’s market share has fallen significantly in some markets. Anthropic, for its part, has established itself as a reference in software engineering and has become one of the preferred tools among programmers. In this race to set the pace of AI, this Thursday we witnessed a curious movement: the almost simultaneous arrival of two models focused on programming, GPT-5.3-Codex and Claude Opus 4.6. The coincidence does not seem coincidental and reflects the extent to which the major players in the sector compete to define the next step, in a scenario where the main beneficiaries are, once again, the users. With these new models already on the table, the question becomes what they really contribute. There are plenty of promises and they are also beginning to appear benchmarks comparable that help to place them. So, therefore, it is time to look in a little more detail at what OpenAI and Anthropic propose for those who use AI as a development tool. GPT-5.3-Codex and Opus 4.6 enter the scene: what each promises to developers GPT-5.3-Codex is presented as a model focused on scheduling agents which seeks to expand the scope of what a developer can delegate to AI. OpenAI claims that it combines improvements in code performance, reasoning and professional knowledge over previous generations and is 25% faster. With this balance, the system is oriented to prolonged tasks that involve research, use of tools and complex execution, while also maintaining the possibility of intervening and guiding the process in real time without losing the work thread. One of the most striking elements that OpenAI highlights in this generation is the role that Codex itself would have had in its development. The team used early versions of the model to debug training, manage deployment, and analyze test and evaluation results, an approach that accelerated research and engineering cycles. Beyond that internal process, GPT-5.3-Codex also shows progress in practical tasks such as the autonomous creation of web applications and games. The company has published two examples that we can try right now by clicking on the links: a racing game with eight maps and a diving game to explore reefs. Anthropic’s turn comes with Claude Opus 4.6, an update that the company presents as a direct improvement in planning, autonomy and reliability within large code bases. The model, they claim, can sustain agentic tasks for longer, reviewing and debugging its own work more accurately. The idea is that we can use these capabilities in tasks such as financial analysis, documentary research or creating presentations. Added to this is a context window of up to one million tokens in beta phase, a leap that seeks to reduce the loss of information in long processes and reinforce the usefulness of the system. Beyond the core of the model, Anthropic accompanies Opus 4.6 with a series of changes aimed at prolonging its usefulness in real workflows. Among them there are mechanisms such as the so-called “adaptive thinking”, which allows the system automatically adjust the depth of your reasoning depending on the context. Configurable effort levels and context compression techniques designed to sustain long conversations and tasks without exhausting the available limits also appear on the scene. Added to this are teams of agents that can be coordinated in parallel within Claude Code and deeper Excel or PowerPoint integration. While OpenAI’s product, GPT-5.3-Codex, is not yet available in the API, Anthropic’s is. Maintains the base price of $5 per million entry tokens and $25 per million exit tokenswith nuances such as a premium cost when the prompts exceed 200,000 tokens. Measure who wins with numbers? When trying to put GPT-5.3-Codex and Claude Opus 4.6 face to face, the main obstacle is not the lack of figures, but rather their difficult correspondence. Each company selects evaluations that best reflect its progress and, although many belong to similar categories, they differ in methodology, versions or metrics, which prevents a direct reading. In this type of models, this fragmentation of results is part of the state of the technology itself, but also requires cautious interpretation that separates technical demonstrations from truly equivalent comparisons. Only from this filter is it possible to identify the few points where both systems can be measured under comparable conditions and draw useful conclusions for developers. If we restrict the analysis to truly comparable metrics, the common ground between GPT-5.3-Codex and Claude Opus 4.6 is limited to two specific evaluations identified through our own research: Terminal-Bench 2.0 and OS World in its verified version. The results show a distribution of strengths rather than a clear supremacy. GPT-5.3-Codex marks a 77.3% in Terminal-Bench 2.0 compared to 65.4% for Opus 4.6, which points to greater efficiency in terminal-centric workflows. On the contrary, Opus 4.6 reaches a 72.7% on OSWorldsurpassing the 64.7% of GPT-5.3-Codex in general interaction tasks with the system, a contrast that reinforces the idea of ​​specialization according to the environment of use. So we could say that the capabilities described by each manufacturer point to tools that are no longer limited to generating code, but rather seek to participate in prolonged processes of analysis, execution and review within real professional environments. This transition introduces new selection criteria that go beyond punctual performance. In Xataka | OpenAI has a problem: Anthropic is succeeding right where the most money is at stake

Anthropic has taken Apple’s strategy against Microsoft to the Super Bowl: making using the rival look ridiculous

by

Anthropic has opened the Super Bowl by attacking OpenAI with ads that show virtual therapists advertising dating apps and personal trainers selling boosts for short people. The message: “Ads are coming to AI. But not to Claude“(“The ads are reaching the AI. But not Claude.”) Sam Altman has responded in X calling them “dishonest” and accusing them of “doublespeak“, “double speech” in Spanish, although a better adapted translation could be “deceptive language” or simply “hypocrisy.” It seems like a minor skirmish, two rivals fighting over an advertisement. But under that hood is a billion-dollar question: What kind of business will AI be when it’s established? The history of the Internet is summarized in two great models: One free supported by advertising: Google, Facebook, YouTube, Instagram, TikTok… regardless of whether they have premium versions. Other direct payment by subscription: Netflix, DAZN, Disney+, Apple Music, PSN… The first aims to maximize the audience, the second aims to maximize the revenue per user. The AI ​​is right now deciding which of the two paths it takes. In Xataka AI is breaking one of the oldest economic paradigms in history: that cheap equals "bad" OpenAI has already chosen and is starting to test putting ads on free ChatGPT accounts. Altman justifies it with the classic argument of democratization: “More Texans use free ChatGPT than the total number of people using Claude in the United States.” In other words: they want to reach those billions of people who are not going to pay 20 dollars a month. And for that you need advertising. Anthropic chooses the opposite. “Anthropic offers an expensive product to rich people,” Altman reproaches him. In a way, it is true: Claude is betting above all on contracts with companies and premium subscriptions of 20, 100 and 200 dollars per month. Their model depends on the AI ​​being valuable enough for you to pay for it. And so that you look from time to time to the higher plan with the temptation to go up one more step. Without advertising, without sponsored links and without responses being influenced by advertisers. The difference is not only business, it is product. An AI with advertising has different incentives than one without it. What happens when you ask the assistant what car to buy you and there is a manufacturer paying to appear in their answers? What about medical, financial, legal advice? OpenAI has promised that “ads do not influence responses.” That’s what he said in minute 0. But that promise will be increasingly difficult to sustain as monetization pressure increases. {“videoId”:”x9u4ml2″,”autoplay”:false,”title”:”Does Gemini 3 surpass ChatGPT? This is Google’s new AI”, “tag”:”Webedia-prod”, “duration”:”156″} Anthropic has its own problem: If it only reaches those who can afford to pay, AI becomes a tool of the elites. A technology that promises to democratize knowledge ends up reproducing the class divisions that already exist. We saw this coming with the arrival of $200 plans to access the AI ​​elite. A gap that creates another gap, The parallel with the history of the Internet is inevitable. Free social networks caught (almost) all of us in the 1910s, but in return they built advertising surveillance machines optimized for the engagementnot for anyone’s well-being. Payment services are cleaner, but also more exclusive. So AI is now at that bifurcation point: OpenAI is committed to being the YouTube of AI: free for everyone, supported by ads and with premium versions for those who want to pay. Anthropic wants to be the Netflix: better experience and free of ads, but only for those who pay. It is true that it maintains a free plan, but its limits are a continuous invitation to check out or leave. And now it’s up for grabs What kind of relationship with those machines that know more and more about us and from which we ask more and more?. Whether they will be services that serve us or whether they will be platforms that monetize us. In Xataka | The AI ​​of 2026 brings an uncomfortable truth: the most useful will be the one that watches us the most Featured image | Anthropic (function() { window._JS_MODULES = window._JS_MODULES || {}; var headElement = document.getElementsByTagName(‘head’)(0); if (_JS_MODULES.instagram) { var instagramScript = document.createElement(‘script’); instagramScript.src=”https://platform.instagram.com/en_US/embeds.js”; instagramScript.async = true; instagramScript.defer = true; headElement.appendChild(instagramScript); – The news Anthropic has taken Apple’s strategy against Microsoft to the Super Bowl: making using the rival look ridiculous was originally published in Xataka by Javier Lacort .

Anthropic has rewritten his 25,000-word “Constitution” for Claude. It is the manual for how AI should behave

by

Anthropic has published a completely renewed version of the so-called “Claude Constitution”. Yes friends, an AI also needs a constitution, or at least a series of documents that explain with total transparency what direction the company has decided to take with its AI tool. It is a way to save us trouble in the event that become aware. The document The question in question consists of 80 pages and nearly 25,000 words, and basically shows what values ​​Anthropic relies on to train its models and what they hope to achieve with it. Alluding to Asimov, it would be something like a broader and more complex version of his three laws of robotics. Why it is important. Anthropic carries a good time trying to differentiate from OpenAI, Google or xAI, wanting to position itself as the most ethical and safe alternative on the market. This Constitution is the centerpiece of their training method called “Constitutional AI”, where the model itself uses these principles to self-criticize and correct its responses during learning, instead of relying exclusively on human feedback. The document is not written for users or researchers: it is written for Claude. It was time to update. The first version of the Constitution, published in 2023, was a list of principles drawn from sources such as the UN Universal Declaration of Human Rights or, as they mention from Fortune, from Apple’s terms of service. Now, according to Anthropic, they have taken a completely different approach: “To be good actors in the world, AI models like Claude need to understand why we want them to behave in certain ways, rather than simply specifying what we want them to do,” affirms the company in its statement. The new document is structured around four fundamental values, and the most interesting thing is that Claude must prioritize them in this order when they conflict: Be largely secure: Do not undermine human AI oversight mechanisms during this critical phase of development. Be broadly ethical: act honestly, according to good values, avoiding inappropriate, dangerous or harmful actions. Comply with Anthropic guidelines– Follow specific company instructions when relevant. Be genuinely helpful: benefit the operators and users with whom it interacts. The majority of the document is concerned with developing these principles in more detail. In the utility section, Anthropic describe to Claude as “a brilliant friend who also possesses the knowledge of a doctor, lawyer and financial advisor.” But it also sets absolute limits, called “hard constraints,” that Claude must never cross: not provide significant assistance for bioweapon attacks, not create malware that can cause serious harm, not assist in attacks on critical infrastructure such as power grids or financial systems, and not help “kill or incapacitate the vast majority of humanity,” among others. Consciousness. The most striking part of the document appears in the section titled “The Nature of Claude,” where Anthropic openly acknowledges its uncertainty about whether Claude could have “some kind of conscience or moral status.” “We are concerned about Claude’s psychological safety, sense of identity, and well-being, both for Claude’s own sake and because these qualities may influence his integrity, judgment, and safety,” they count from the company. The company claims to have an internal team dedicated to “model well-being” that examines whether advanced systems could be sentient. Amanda Askell, the Anthropic philosopher who led the development of this new Constitution, explained told The Verge that the company doesn’t want to be “completely dismissive” about this issue, because “people wouldn’t take it seriously either if you just said ‘we’re not even open to this, we don’t investigate it, we don’t think about it.’” The document also raises complex moral dilemmas for Claude. For example, it states that “just as a human soldier might refuse to shoot peaceful protesters, or an employee might refuse to violate antitrust law, Claude should refuse to assist with actions that concentrate power in illegitimate ways. This is true even if the request comes from Anthropic itself.” And now what. Anthropic has published the entire Constitution under a Creative Commons CC0 1.0 license, meaning anyone can freely use it without asking permission. The company promises to maintain an updated version on its website, considering it to be a “living document and a continuous work in progress.” Cover image | Andrea De Santis and Anthropic In Xataka | Company CEOs say AI is saving them a day of work a week. Employees say otherwise

The AI ​​Claude Code “only” programmed. With Cowork, Anthropic wants its AI to take care of everything else

by

Claude Code has become a revolution for programmers, but at Anthropic they are not satisfied with that, and now they want their Claude family AI models to serve much more. And that’s why have created Coworka different agent, especially ambitious and who opens the door to fantastic options… if you trust him. What is Cowork. Those responsible for this project have taken the foundations of Claude Code and applied them to the Claude desktop application (for now, only the macOS one). But they have also done something equally special: giving Claude permission to access a specific folder on our computer and, from there, he can take control of those files and work with them as we want. Hello, robot-secretary. Instead of access to the “vibe coding” we will have access to a kind of “vibe working”. Thus, we can ask Cowork to do all kinds of operations with those files: If we have a folder full of disorganized icons, we can ask you to ordered them to us and reorganize them all into folders by file type or theme If there are a lot of photos of receipts in that folder, we can tell you to create an expense report If what we have is a bunch of digital voice or text notes, we can ask you to write a report summarizing and combining them all. If we have a folder full of podcasts, we can have it go through it, analyze it and summarize the top 10 points of all of them or transcribe them If you have all your financial trading and investment reports and data, you can ask them to create a final report for you. help you declare them If you have videos and want to find one of a squirrel and then convert it to another format, also does. Full autonomy. We are therefore faced with an AI agent capable of accessing our files, analyzing them and working with them to generate new information and useful content from all that data. And we only have to ask it with natural language, because the agent is capable of understanding it, asking us questions if it needs more details, and then solving the task autonomously even if it involves several steps. Cowork operates in a container. The way CoWork works allows you to grant permission to certain folders, but when the AI ​​operates on said files it does so in isolation. As explains Simon WillinsonClaude uses a virtual machine and downloads and boots a custom Linux file system to operate on those files independently and isolated, which theoretically guarantees that our files are theoretically safe and Cowork does not access anything that we have not given permission to. Connections to other apps. In addition to being able to work directly with your files, Cowork benefits from its ability to connect with other applications that you have installed on your computer. You can use ffmpeg to convert the squirrel video, Asana if you want to organize your notes into projects, or an office application if you need to create a spreadsheet. But we will have to trust. Willinson himself warns that these types of systems have the danger of someone “hacking” them with jailbreaking or prompt injection techniques that now become more dangerous because, as we say, what Cowork does is work on our files. And of course we have to be careful with the information and data we share with CoWork: those responsible for Anthropic themselves have a document to “use it safely“. Limited release. Cowork is available as a “research preview”, and is only available to users of the Claude Max subscription which costs between $100 and $200 per month. It is clear that at Anthropic they prefer to go step by step with a very powerful but also delicate feature if we do not use it with caution: in the end we are giving access to our files to an AI, and we know that AIs can make mistakes. An AI on your computer. This release from Anthropic points to what all AI agents that want to conquer our computer should theoretically point to. Since that Computer Use that Anthropic launched in October 2024, things have come a long way, and little by little we are getting closer to that future in which we will be able to work with our computer in a very different way than we did until now… if we want and trust AI, of course. In Xataka | Operator also “looks” at the screen and moves your mouse for you like other AI agents. It does it better thanks to CUA

OpenAI, Google and Anthropic fight among themselves. Samsung fights everyone else elsewhere

by

Samsung has presented at the CES 2026 its “AI philosophy,” a grandiloquent concept that sums up its strategy: using its 430 million SmartThings users as moat (or ‘defensive moat’) against the invasion of AI in homes. Why is it important. OpenAI, Google and company remain focused on announcing the most powerful model. There is little to do against them on that side if you haven’t been doing it for years, so Samsung is playing something else that is not about winning the algorithm war, but about controlling where those algorithms live. SmartThings is not just an app. It is a platform Matter compatible that connects hundreds of millions of devices already in homes around the world. That means Samsung can add AI to products people already use, without asking them to buy anything new or change their habits. Others have to convince you to put a smart speaker in the kitchen. Samsung already has your refrigerator, your television, your washing machine and your vacuum cleaner. And everyone talks to each other. Between the lines. Samsung’s “AI philosophy” seems, above all, a response to Amazon with its Alexa+. Both proposals have things in common: they understand that if AI models tend to commoditize (to be technically equal until they are not easily distinguishable), the value is in who has the speaker in your kitchen, the TV in your living room and the refrigerator that knows what you eat. Samsung has been building that ecosystem for years and now it is activating it for something else. Implementation makes the difference: Family Hubwith AI and Gemini vision, recognizes what you put in and out of the refrigerator, suggests recipes and connects with other appliances. It’s real tracking so that when you ask yourself “what can I make for snack-dinner?”, the system suggests recipes based on what you have, not on an inventory you made by hand three weeks ago. Vision AI Companion It recognizes what you’re watching on TV and suggests recipes if food appears on the screen. Then send that recipe to the Family Hub in your refrigerator, which checks what ingredients you have and tells you what you’re missing. If you decide to cook it, send the instructions to the oven so that it is preheated to the exact temperature. AI Soccer Mode Pro Automatically adjusts image and sound when it detects that you are watching football. You can turn up the audience volume, turn down the commentators, or balance both. It’s AI applied to something as specific as “I want to enhance the field atmosphere” or “I want to prioritize the narrator’s voice.” It is perhaps not as attractive an approach as the war of chatbots that are increasingly capable of more, but maybe (just maybe) it will end up being more profitable. And something else: SmartThings as a Matter-compatible standard. That expands the potential ecosystem far beyond Samsung’s own products. Yes, but. There are two weak points in that strategy: Samsung depends on third-party models. Gemini is your main partner, also for the home, for the smart component. If the models run out commoditizingwe will have to compete on price. And in the price war there always appears a Chinese manufacturer willing to go lower. privacy. An ecosystem that knows what you eat, what you see, when you sleep or how you move is also an ecosystem that can monetize that data. The last threat It’s called Dreame. and there is a red flag On that second point: Samsung has announced an agreement with the insurer HSB to give discounts on home insurance in exchange for connecting home appliances to SmartThings. That is, saving some money in exchange for handing over your behavioral data. As what we already saw with health insurance and wearables. It’s a double-edged sword: if your behavior reduces your premium, it can also increase it. Or directly invalidate coverage. The bet. If it works, Apple will speed up with Home (previously HomeKit), Google will push with its Nest and Amazon will double down with Alexa+ and Ring. The battle is no longer for the best language model. It’s because more devices in more homes capturing more data. Samsung has been losing ground in mobile phones for years fruit of Apple’s clamp in premium and Chinese manufacturers in price. Also against LG in some appliances not to mention Chinese baking for the home. But in the sum of connected devices per home, it does not have so many rivals. That is its trump card: converting the fragmentation of its catalog into the advantage of its ecosystem. The question is whether consumers will give up control of their home in exchange for convenience. The answer determines whether Samsung ends up being the silent winner of the AI ​​era or simply the maker of gadgets that run other people’s intelligence. In Xataka | I would never have imagined answering a call from the washing machine. Until I tried the latest from Samsung Featured image | Screens even in washing machines and appliances that talk to each other: this is how Samsung imagines the future of the connected home

To the question of what sense it makes to compete with Google, OpenAI or Anthropic in AI, Mistral has an answer: small and local models

by

French startup Mistral AI Mistral 3 has been launcheda family of 10 open source artificial intelligence models that represent its most ambitious commitment to date. The Parisian company, which is often considered the main European hope in the development of AI, seeks to differentiate itself from the large American technology companies by betting on flexibility and deployment in all types of devices instead of raw power. Under these lines we tell you all the news. What Mistral has presented. The Mistral 3 family includes a flagship model called Mistral Large 3, with 675 billion parameters, and nine compact models grouped under the name Ministral 3 (in three sizes: 14,000, 8,000 and 3 billion parameters). All models are released under Apache 2.0 license, allowing unrestricted commercial use. The large model also has multimodal capacity, being able to process text and images. It is also multilingual, with a special emphasis on European languages. On the other hand, small models can run on devices with just 4 GB of memory, making them perfect for modest laptops, mobile phones and embedded systems without the need for an internet connection. Why strategy matters. While OpenAI, Google and Anthropic focus on increasingly powerful and closed systems with agentic capabilitiesMistral has focused on the breadth and scope of its models, efficiency and what its co-founder Guillaume Lample calls “distributed intelligence.” According to declared told VentureBeat, the company believes the future of AI is defined not by scale, but by ubiquity: models small enough to run in drones, vehicles, robots and consumer devices. The economic and practical argument. Lample explained It means that in more than 90% of cases, a small, specifically tuned model can get the job done, especially if it is trained with synthetic data for specific tasks. According to Lample, this is not only cheaper and faster, but it eliminates concerns about privacy, latency and reliability. The company also has teams that work directly with customers to analyze specific problems and fine-tune small models that perform specific tasks. This, above all, can attract companies that become frustrated when choosing the best possible model for a specific task and, if it does not perform adequately, they end up giving up. Europe is lagging behind. If we talk about innovation and technology around AI, we do not hesitate to say that Europe is leagues away of what companies in the United States and China are offering. This is why Mistral AI advocates a different approach in which it prioritizes massive deployment in devices and the flexibility of its smaller models. The capacity offered by open models can be a great asset to continue betting on these technologies. In China, for example, the open models of DeepSeek, Alibaba or Kimi are emerging widelyabove in certain tasks even competitors as large as ChatGPT. Lample explained that most leading Chinese models are exclusively text-based, with separate image processing systems. For this reason, they also want to opt for a multimodal approach. A complete ecosystem. Mistral no longer only offers language models. The company has built an entire ecosystem that includes Mistral Agents APIwith connectors for code execution, web search and image generation; Masterlyyour reasoning model; Mistral Code for programming assistance; and AI Studioan application deployment platform that also has analytical and logging capabilities. Furthermore, his assistant Le Chat It has incorporated a deep research mode, voice capabilities and a list of more than 20 enterprise integrations. Thus, in addition to its model offering, the company can provide other companies with a whole layer of personalized products and services, with the aim of being their main source of financing. Digital sovereignty. Although Mistral is often characterized as Europe’s answer to OpenAI, the company prefers to consider itself as ‘a transatlantic collaboration’. Its CEO, in fact, is in the United States, has teams on both continents and trains these models in collaboration with American teams and infrastructure. However, its positioning as a defender of European digital sovereignty has earned it strategic partnerships with the French army, the country’s employment agency, the Luxembourg government and various European public organizations. The European Commission presented in October a strategy to promote European AI tools that provide security and resilience while boosting the continent’s industrial competitiveness. Offline capabilities for democratization. The use cases that Mistral has designed for its small models include, above all, local applications, such as factory robots that use sensor data in real time and without relying on the cloud, drones in natural disasters or rescues that operate offline, and smart cars with functional AI assistants in remote areas. Lample stood out that there are billions of people without internet access but with laptops or cell phones capable of running these small models, which he considers potentially revolutionary. Additionally, by running on the device, these apps preserve the privacy of user data. Real “open source” debate. Not everyone celebrates Mistral’s approach. Some critics question his decision to opt for models’open weight‘, that is, free to access but providing less information about their code than truly “open source” models, which provide the code and training data necessary to train a model from scratch. Andreas Liesenfeld, assistant professor at Radboud University and co-founder of the European Open Source AI Index, declared to the Financial Times that data at scale is the missing key in the European AI innovation ecosystem and that Mistral does not contribute to that at all. The long-term strategic bet. Lample recognize that their models are “a little behind” the most advanced closed systems, but argued that the important thing is that “they are catching up quickly.” Time will tell if Mistral’s approach to low-cost, versatile models with local applications ends up working for them to end up positioning themselves as one of the great European bets on AI. Cover image | Mistral AI In Xataka | China already has an army of 5.8 million engineers. His new plan involves accelerating doctorates

Log In

Forgot password?

Forgot password?

Enter your account data and we will send you a link to reset your password.

Your password reset link appears to be invalid or expired.

Log in

Privacy Policy

Add to Collection

No Collections

Here you'll find all collections you've created before.