The AI revolution in programming has shaken journalists, data scientists, and economists, but its epicenter has been among developers. Andrej Karpathy, one of the most influential software engineers, expressed it this way: “It’s hard to convey how much programming has changed due to AI in the last two months: it hasn’t been gradual, but a sudden leap since December. Before, code agents didn’t work; now they do.” Karpathy, who used to write 80% of his code, now hardly programs. The driver of this change is Claude Code. Let’s analyze what has happened and why Anthropic, the company behind Claude, is at the center of the conversation. Then, we’ll see what Claude is for and how I use it. Until recently, AIs were mainly conversational: they answered questions. But they have evolved. Chatbots like ChatGPT, Gemini, and Claude have gained autonomy: they search for information, summarize it, and reason step by step. The models of 2022 were chatterboxes; those of 2026 execute tasks. This is evident in the free versions, but especially in the paid agents: they open files, write and correct code, and execute it. Claude Code, Anthropic's agent, has driven this leap. Programmers have moved from using AI as an advanced autocompleter to delegating complete tasks: you give instructions, and the agent works for minutes, generating files, thousands of lines of code, functional websites, and calculations. The agents operate in cycles of trial and error, solving problems autonomously. They correct themselves: “my attempt failed; I check the directory and try again.” They are not infallible and require supervision, but they work without constant intervention. For those of us who program, the change is evident. The impact is also reflected in Anthropic's figures. Its CEO, Dario Amodei, describes its growth as “bizarre”: in 2023, from zero to 100 million; in 2024, to one billion; in 2025, to 10 billion. By March 2026, Anthropic would have reached 19 billion dollars in annualized revenue. According to Epoch AI, it could surpass OpenAI, creators of ChatGPT. The market confirms this trend. Menlo Ventures estimates that Anthropic's share of the enterprise AI market grew from 12% to 40% between 2023 and 2025. In programming, Claude leads with a 54% market share. Claude offers three modes: Chat, Code, and Cowork. Chat is for conversation and document analysis. Code is for programming. Cowork, the most recent, is a versatile agent that manages files, works in Excel, browses, and executes tasks. Chat is free; Code and Cowork require a subscription (about 18 euros per month). I have been using Claude for 18 months. For the past year, it has been my main AI. Some examples: 🏠 I use Cowork at home. When we moved, I created a folder with appliances. I indicated the models, and Claude searched for and downloaded the manuals. Now I ask it questions: “How do I set up the wifi for the air conditioner? Does the dryer have a setting for sheets?” 🧠 I use it to learn. I ask it about transformer architecture or intelligence in elite football players. It’s not Google: it’s a conversation adaptable to my level. I can ask for papers to delve deeper. 📄 I automate administrative tasks. I created a skill for invoices: Claude extracts data from emails, uses a template, and generates the PDF. I just review and send. 📊 It helps me with electoral prediction. Claude understands and adapts my code, reviews texts, and verifies figures with other sources. As a quantitative editor, it is unbeatable. 🔍 Claude programmed our visual explorer for 23F. Without AI, that website wouldn’t exist. 📬 Cowork improves this newsletter. Every Thursday, it reviews my sources and reminds me of ideas. It searched for data, located quotes, and edited the text, eliminating redundancies. I’m not the only one. Claude is used at NASA and in international conflicts. It also powers user-created websites for career choices, tax management, train inquiries, or monitoring public contracts. Projects that used to take days or weeks now exist thanks to AI. Anthropic also benefits. In December, they stated that Claude writes between 70% and 90% of its code. The head of Code hasn’t programmed manually for months. In 52 days, they launched 74 updates, as if one AI were building another AI. Claude can help any computer user. It’s a good time to start, but don’t expect miracles: agents don’t guess your intentions; they require clear instructions, knowledge of their limits, and supervision. It’s like delegating to a person: with caution. This points to the future. Jensen Huang distinguishes between tasks and purpose in work. Tasks can be automated, but purpose remains. For a software engineer, the goal is not to write code but to create useful applications. Huang believes programming will be about “specifying” what is wanted, and the number of programmers will grow from 30 million to over 1 billion. The programmer still decides which problem to solve and how the solution should be. Their creativity is key. The same goes for journalism or analysis: human judgment distinguishes the excellent from the mediocre. And the responsibility to validate the result remains ours, not Claude’s. Source: elpais.com