Cursor 2.0 shipped Saturday, six weeks after the Windsurf merger closed and the combined company had to decide what it actually was. The answer, judging by the release, is “team orchestration layer with an editor attached,” which is roughly the inverse of the 2024 pitch and a much harder product to build. The headline feature is background agents that run on Cursor’s hosted compute, take a written brief from a human, file a PR, and tag a reviewer. Each agent gets routed to a model based on the task type. Claude Sonnet 4.6 for refactors, GPT-5.5 for debugging, the merged Cursor-Windsurf in-house model (now branded Compass) for repo-wide changes. The routing layer is the new product surface.
The pricing is also the new product surface. Cursor 2.0 introduces seat tiers with included agent-hours, on top of the existing per-completion pricing. The implied bet is that a team of ten engineers using Cursor will be running thirty or forty background agents at any given time by 2027, and the right unit to price is agent-hours, not autocompletes. Enterprise pricing was not disclosed. The Anysphere blog post used the phrase “platform deal” twice, which is enterprise sales code for “we will negotiate.”
The product question this raises is the one Cognition has been raising since Devin shipped. If the agents are running on the vendor’s hosted compute and filing PRs against your repo, what exactly is the IDE for. Cursor’s bet is that the IDE is the surface where humans review, edit, and merge agent work, and the actual writing of new code by humans is increasingly a fallback. That is consistent with where GitHub Copilot has been heading since the autonomous-agent preview last quarter. It is also consistent with the procurement conversation every enterprise engineering leader has been having since January, which is “how do I bill for AI engineering work that has no human keystrokes attached to it.”
The bigger read is that the editor wars are over. The orchestration wars just started, and Cursor 2.0 is the first product to admit that out loud.