Cursor Unveils Agent-First Coding Platform Cursor 3 to Compete With OpenAI and Anthropic
On Thursday, AI code developer Cursor announced the launch of Cursor 3, a new product interface that enables users to spin up autonomous AI coding agents to complete end-to-end development tasks on their behalf. Built under the codename Glass, Cursor 3 marks the startup’s direct response to the breakout popularity of agentic coding tools from leading AI labs, including Anthropic’s Claude Code and OpenAI’s Codex, which have gained millions of developer users in recent months.
“In the last few months, our profession has completely changed,” Jonas Nelle, one of Cursor’s co-heads of engineering, told WIRED in an interview. “A lot of the product features that got Cursor to where we are today just aren’t as important moving forward.”
Cursor increasingly finds itself competing with the world’s top AI labs for both individual developer users and enterprise clients. The company pioneered one of the first and most widely adopted AI-assisted coding workflows, integrating models from OpenAI, Anthropic, and Google directly into its editor to become one of the largest AI customers for all three firms. But over the past 18 months, OpenAI and Anthropic have launched their own native agentic coding products, priced via heavily subsidized subscriptions that have put significant financial pressure on Cursor’s independent business.
Where Cursor’s core original product offered developers an AI-augmented integrated development environment (IDE) that let users call on AI for incremental help, new tools like Claude Code and Codex are built around letting developers offload entire tasks to autonomous AI agents—even supporting multiple parallel agents for large, complex projects. Cursor 3 is the startup’s take on a fully “agent-first” coding product. Per Nelle, the platform is optimized for a future where developers spend their days “conversing with different agents, checking in on them, and reviewing the work that they did,” rather than writing every line of code manually.
Cursor 3’s new agent interface launches directly within the company’s existing desktop app, running alongside the traditional IDE. At the center of the new agent workspace is a plain-text input box, where users can describe their desired task in natural language, giving the experience a chatbot-like feel far removed from a traditional coding environment. After a user submits their prompt, the AI agent begins work immediately, no manual code input required. A left-hand sidebar lets developers easily view and organize all active AI agents running in their Cursor session.
Compared to the standalone desktop apps for Claude Code and Codex, Cursor 3’s key differentiator is its deep integration between the agent-first workflow and Cursor’s existing AI-powered development environment. During a demo for WIRED, Alexi Robbins, Cursor’s other co-head of engineering leading the Cursor 3 project, showed how users can prompt a cloud-based agent to build a full new feature, then seamlessly review and edit the AI-generated code locally on their own machine through Cursor’s native editor. Nelle and Robbins note that the company does not prioritize which interface developers use most—their core goal is simply keeping users within the Cursor ecosystem.
Competing With the Big AI Labs
Last week, I visited Cursor’s San Francisco headquarters in the city’s North Beach neighborhood, where the fast-growing startup has recently expanded into a converted old movie theater. Rumored to be raising new capital at a $50 billion valuation—nearly double its valuation from a funding round last fall—the company’s growth is visible in small, telling details: where employees once tossed their shoes in a loose pile by the front door, there is now a row of large, organized shoe racks, a quiet sign of the startup maturing.
Even so, Cursor retains the scrappy culture of an early-stage startup, a trait employees say is a core draw of working there: the team can ship product updates quickly and avoids the bureaucratic bloat of large tech firms. But as the company races to catch up to Anthropic and OpenAI in the agentic coding race, that agility alone may not be enough. The battle to build the world’s leading AI coding agent is shaping up to be Cursor’s most capital-intensive chapter yet.
Multiple developers told WIRED they have shifted the majority of their AI coding work away from Cursor to Claude Code and Codex, with subsidized pricing as the top reason. WIRED has previously reported that Claude Code and Codex subscribers can access well over $1,000 worth of model usage for their $200 monthly plans, a value Cursor cannot easily match.
Ronald Mannak, founder of Pico AI, a startup building AI tools for Apple developers, told WIRED he has mostly moved away from using Cursor and rival Windsurf to agent-first products like Claude Code and Codex. His decision, he says, is driven almost entirely by which tool offers the most generous usage rate limits. Jack Crawford, co-founder of AI memory startup mVara, says he rarely uses Cursor or Windsurf today, despite relying heavily on both tools last year. He now uses Claude Code full-time for the unbeatable subscription value.
Cursor offered heavily subsidized subscription plans for its AI coding tool until June 2025, when the startup announced it would transition to usage-based pricing for developers. The change upset many users at the time, but it was a deliberate move to improve margins and build a more sustainable long-term business for the independent startup. OpenAI and Anthropic have raised tens of billions of dollars more in capital than Cursor, so they can afford to spend heavily on customer acquisition via subsidized pricing (though Anthropic has already begun adjusting rate limits for Claude Code subscriptions). Still, Cursor says it has alternative strategies to compete with the leading AI labs.
One key strategy is training custom in-house AI models that the company can serve to customers at low cost. The startup recently launched Composer 2, an in-house model built on an open-source base from Chinese AI lab Moonshot AI, with additional pre-training and post-training completed by Cursor’s team. Nelle says developers typically choose AI models based on a mix of performance, price, and speed—and he argues Composer 2 is competitive on all three fronts. Cursor also plans to train future Composer model generations entirely from scratch.
That said, training large AI models is an extremely expensive undertaking. While Cursor has a history of delivering strong results with limited resources, the global AI coding race is heating up rapidly. OpenAI and Anthropic have both recognized the enormous market opportunity for AI coding tools and are investing heavily in the space. All major players in the space are converging on similar agent-first roadmaps, where AI agents take on an increasingly large share of a developer’s daily workload. In this new agent-first landscape, it remains unclear how Cursor can stay competitive without raising significantly more capital—and quickly.
This is an edition of Maxwell Zeff’s Model Behavior newsletter. Read previous newsletters here.
