Cursor 2.0 Released: In-House Composer Model Debuts, Pivot to Multi-Agent Coding

Cursor released version 2.0 on October 29, 2025. The headline is Composer, its first in-house frontier model — an MoE model purpose-built for low-latency agentic coding that the company says is 4x faster than similarly intelligent models.

Cursor 2.0 also pivots the product toward multi-agent work: a new multi-agent interface lets developers run several coding agents in parallel on different tasks while monitoring progress in real time — a shift from single-session assistance to multi-task orchestration.

From Wrapper to Model Maker: An Identity Change

Composer means far more than a new model: Cursor long wore the "Claude/GPT wrapper" label, paying massive API bills to vendors who were simultaneously building competing gateways with Claude Code and Codex — supplier-as-competitor is not a sustainable position. An in-house low-latency MoE model claws back cost, experience and bargaining power, and validates our annual review's call: the model layer and tool layer are merging, and the pure-wrapper window has closed.

The Multi-Agent Interface: An Interaction Watershed

The other half of 2.0 is interaction: from question-and-answer pair programming to conducting several agents at once, the developer's role shifts from operator to dispatcher. It rhymes with Claude Code's subagents and Codex's parallel cloud tasks (see our coverage): once long-horizon autonomy stretches, single-threaded interaction becomes the bottleneck and parallel orchestration is the inevitable answer.

The Field: A Leader Under Siege

Cursor tops the AI-native IDE category while surrounded on every side: GitHub Copilot leans on incumbent distribution, Windsurf landed at Cognition after its acquisition saga, ByteDance's Trae attacks on price, and vendor tools Claude Code and Codex flank from the terminal (see our coding-tools annual review). The in-house model is Cursor's key move to raise the moat under siege.

Our Take

Cursor followed up in 2026 with Composer 2 at $0.50/M input tokens, doubling down on cost efficiency. The sample lesson of Cursor 2.0: application companies building models in reverse is no longer fantasy — with real usage data and a scoped objective, a small specialized model can beat general flagships in its lane. The boundary between tool companies and model companies blurred for good that day.