커서, 클로드 코드·코덱스 맞서는 신규 AI 코딩 에이전트 출시

Comment Loader Save Story Save this story Comment Loader Save Story Save this story Cursor announced Thursday the launch of Cursor 3, a new product interface that allows users to spin up AI coding agents to complete tasks on their behalf. The product, which was developed under the code name Glass, is Cursor’s response to agentic coding tools like Anthropic’s Claude Code and OpenAI’s Codex, which have taken off with millions of developers in recent months. “In the last few months, our profession has completely changed,” said Jonas Nelle, one of Cursor’s heads of engineering, in an interview with WIRED. “A lot of the product that got Cursor here is not as important going forward anymore.” Cursor increasingly finds itself in competition with leading AI labs for developers and enterprise customers. The company pioneered one of the first and most popular ways for developers to code with AI models from OpenAI, Anthropic, and Google—making Cursor one of these companies’ biggest AI customers. But in the last 18 months, OpenAI and Anthropic have launched agentic coding products of their own, and started offering them through highly subsidized subscriptions that have put pressure on Cursor’s business. While Cursor’s core product lets developers code in an integrated development environment (IDE) and tap an AI model for help, new products like Claude Code and Codex center around allowing developers to off-load entire tasks to an AI agent—sometimes spinning up multiple agents at the same time. Cursor 3 is the startup’s version of an “agent-first” coding product. According to Nelle, the product is optimized for a world where developers spend their days “conversing with different agents, checking in on them, and seeing the work that they did,” rather than writing code themselves. Cursor is launching its new agentic coding interface inside its existing desktop app, where it will live alongside the IDE. At the center of a new window in Cursor, there’s a text box where users can type, in natural language, a task they’d like an AI agent to complete—it looks more like a chatbot than a coding environment. Press enter, the AI agent sets to work without requiring the developer to write a single line of code. In a sidebar on the left, developers can view and manage all of the AI agents they have running in Cursor. What’s unique about Cursor 3, compared to desktop apps for Claude Code and Codex, is that it integrates an agent-first product with Cursor’s AI-powered development environment. In a demo, Cursor’s other cohead of engineering for Cursor 3, Alexi Robbins, showed WIRED how users can prompt an agent in the cloud to spin up a feature, and then review the code it generated locally on their computer. Nelle and Robbins argue it doesn’t matter which interface developers are spending their time in—they just want people using Cursor. Competing With the AI Labs I visited Cursor's office in San Francisco's North Beach neighborhood last week. The startup is reportedly raising fresh capital at a $50 billion valuation —nearly double what it was valued in a funding round last fall—and has expanded into an old movie theater. Cursor employees used to toss their shoes in a pile by the door upon entry, but now there's a row of large shoe racks, signaling one way in which the company is growing up. Yet Cursor still feels like a startup. Employees tell me that’s part of the appeal of working there; the company can ship quickly and doesn't feel too corporate. But as it finds itself racing to catch up to Anthropic and OpenAI in the agentic coding race, that scrappiness may not be enough. This battle—the one to create the best AI coding agent—may be Cursor’s most capital-intensive chapter yet. Several developers tell WIRED that they’ve shifted most of their AI coding work to Claude Code and Codex, and away from Cursor. A large reason is the aforementioned subsidized subscriptions. WIRED has previously reported that Claude Code and Codex users can get well over $1000 worth of usage for their $200-a-month plans. Ronald Mannak, founder of the startup Pico AI—which makes AI tools for Apple developers—says he’s largely shifted from using Cursor and Windsurf to agent-first products like Claude Code and Codex. He says his decision is largely driven by whichever tool has the most generous rate limit. Jack Crawford, cofounder of the AI memory startup mVara, says he rarely ever uses Cursor or Windsurf anymore, despite heavily using those tools last year. He now goes to Claude Code because of the value of the subscription. Cursor offered a heavily subsidized subscription plan for its AI coding tool until June 2025, when the startup announced it would start charging developers through usage-based pricing . This upset developers at the time, but was part of an effort for the young startup to improve its margins and build a more sustainable business. OpenAI and Anthropic have raised tens of billions of dollars more than Cursor, so they can afford to keep spending heavily on customer acquisition (though Anthropic is starting to adjust its rate limits for Claude Code subscriptions ). But Cursor says it has other strategies to compete with the leading AI labs. Cursor has also started training in-house AI models that it can cost effectively serve to customers. The startup recently launched Composer 2 , an AI model based on an open-source system from the Chinese AI lab Moonshot AI, that Cursor did additional pretraining and post-training on. Nelle tells me people usually pick AI models in Cursor based on some combination of performance, price, and speed—and he argues that Composer 2 is competitive on those fronts. Cursor says it plans to train future Composer models completely from scratch. But training AI models is quite an expensive undertaking. Cursor has historically done well doing more with less, though the AI coding race is now heating up. OpenAI and Anthropic have recognized how large the business around these tools could be and are investing in them heavily. A lot of these companies are also converging on similar products, in which agents are taking on more and more of a developer’s workload. In the agent-first world, it’s hard to imagine how Cursor can stay competitive without raising significantly more capital—and fast. This is an edition of Maxwell Zeff’s Model Behavior newsletter . Read previous newsletters here.