Cursor 2.0 and Composer

Cursor 2.0 (October 2025) shipped Composer, Cursor's first in-house frontier coding model, plus a multi-agent interface for running several agents in parallel. Composer targets sub-30-second agentic turns.

What changed in 2.0

Cursor 2.0 reframes the editor around agents rather than files. Two headline changes: Source: Cursor blog, https://cursor.com/blog/2-0, 2025-10-29

• Composer - Cursor's first coding model, a frontier-level agentic model built for low-latency work inside Cursor.
• A multi-agent interface - a new surface for running many agents in parallel, moving the default workflow away from a single file-by-file loop toward orchestrating concurrent agents.

Composer the model

Composer is positioned as a frontier model that is roughly 4x faster than similarly intelligent models, completing most turns in under 30 seconds. Cursor reports it was trained with reinforcement learning to learn optimal tool-usage patterns from real engineering workflows, and equipped with capable tools including codebase-wide semantic search, which makes it stronger at navigating and editing large repositories. The design bet is that fast, trustworthy multi-step iteration beats marginally higher single-shot benchmark scores for day-to-day coding. Source: Cursor blog, https://cursor.com/blog/2-0, 2025-10-29 Source: Digital Applied, https://www.digitalapplied.com/blog/cursor-2-composer-ai-coding-agent-guide, 2025-10-28

Composer 2.5 (May 2026)

Announced 2026-05-18 as Cursor's "most capable model yet" — more intelligent, better at sustained long-running tasks, and more reliable at following complex instructions; Cursor doubled the model's included usage for launch week. Source: X/@cursor_ai, 2026-05-18

Technical claims from the launch thread: Source: X/@cursor_ai thread, 2026-05-18

• Up to 10x more efficient than similarly capable models while staying "exceptionally intelligent."
• Improved by scaling training, generating more complex RL environments, and new learning methods — notably text feedback during RL to assign credit across rollouts spanning hundreds of thousands of tokens.
• Built on the same open-source base as Composer 2: Moonshot's Kimi K2.5.
• A significantly larger from-scratch model is training with SpaceXAI, using ~10x more total compute on Colossus 2's "million H100-equivalents" — projected as a major capability leap.

So Composer 2.5 is an RL-tuned model on an open base today, with a frontier from-scratch successor in the pipeline. It is exposed as a fast option (composer-2.5-fast) in Cursor's agent model picker. Source: compiled, 2026-06-12

The local visual artifact is a benchmark table, useful mainly for routing expectation rather than absolute model ranking. Cursor's shown numbers: Composer 2.5 at 69.3% Terminal-Bench 2.0, 79.8% SWE-Bench Multilingual, and 63.2% CursorBench v3.1 harder tasks. That puts it near Opus 4.7 on the two Cursor-shown coding rows and above Composer 2, while the image notes that Opus 4.7 and GPT-5.5 use self-reported public-eval scores. Source: X/@cursor_ai image artifact, reviewed 2026-06-30

The June 2026 SpaceX acquisition signal makes the Composer line a vertical-integration story, not only a model release story. The source claim says SpaceX exercised an option to acquire Cursor/Anysphere and that the joint SpaceXAI/Cursor model will ship in Cursor and Grok Build. Treat that as a pending ownership/platform-neutrality signal unless later official sources show the close mechanics; the already-public model-training relationship is confirmed by Cursor's SpaceX training blog and the Composer 2.5 post. Source: X/@SpaceX, 2026-06-16; Source: Cursor SpaceX model-training blog, 2026

Where it fits Kevin's stack

Cursor is Kevin's primary IDE, so Composer is the default agent loop and the multi-agent interface is the native version of fan-out work otherwise done through the Cursor Orchestrate Skill skill (recursive agents via the Cursor SDK). Composer competes head-to-head with GPT-5-Codex (OpenAI's agentic coding model) and the Windsurf Editor line. Context-cost discipline still applies inside Cursor 2.0 - see Lean Ctx for trimming token load at repo bootstrap. For non-Cursor harness comparison, see Everything Claude Code - Agent Harness Performance System. Cursor 3 later added Cursor Design Mode for visual point-and-edit UI work in the Agents Window browser.

---

Timeline

• 2025-10-29 | Cursor 2.0 released with Composer (first Cursor coding model, ~4x faster than comparable models, sub-30s turns, RL-trained, codebase-wide semantic search) and a parallel multi-agent interface. Source: Cursor blog, https://cursor.com/blog/2-0, 2025-10-29
• 2026-05-18 | Composer 2.5 announced as Cursor's most capable model, better at long-running tasks, up to 10x more efficient, built on Moonshot Kimi K2.5; a larger from-scratch model is training with SpaceXAI on ~10x compute (Colossus 2). 13,253 likes. Source: X/@cursor_ai, 2026-05-18
• 2026-05-31 | Tool page created. Cursor's ARR had reached roughly $500M annualized during the 2025 AI-IDE race. Source: TechCrunch, https://techcrunch.com/2025/07/14/cognition-maker-of-the-ai-coding-agent-devin-acquires-windsurf/, 2025-07-14
• 2026-06-16 | @SpaceX announced it had exercised an option to acquire Cursor (Anysphere) in an all-stock transaction; the SpaceXAI-trained joint model (foreshadowed in the Composer 2.5 note) is to ship in Cursor and Grok Build. Same-day Cursor launched Cursor Origin (git hosting for teams and agents). Source: X/@SpaceX, 2026-06-16
• 2026-06-30 | Deep-reviewed the local benchmark artifact and recorded the displayed Composer 2.5 numbers: 69.3% Terminal-Bench 2.0, 79.8% SWE-Bench Multilingual, 63.2% CursorBench v3.1 harder tasks. Source: X/@cursor_ai image artifact, 2026-05-18