On May 18, Cursor released Composer 2.5, its third-generation in-house coding model. For the first time, it stands alongside Opus 4.7 and GPT-5.5 on public benchmarks—while costing only a tenth per task.
Numbers on the board
On CursorBench v3.1 default settings, Composer 2.5 scored 63.2%, versus GPT-5.5's 59.2%. On SWE-Bench Multilingual, Composer 2.5 achieved 79.8%, surpassing GPT-5.5's 77.8%.
However, on Terminal-Bench 2.0, GPT-5.5 still leads with 82.7% to 69.3%—Cursor openly published this result, a rare move among Chinese AI model makers.
Pricing is the real highlight:
- Standard Composer 2.5: $0.50 input / $2.50 output per 1M tokens
- Fast variant (default): $3.00 input / $15.00 output per 1M tokens
"Composer 2.5 is exceptionally intelligent and up to 10x more efficient than similarly capable models." — Cursor
In plain terms: a single task costs under $1, while competitors charge $11. That gap is enough for enterprise CTOs to redo their budgets.
Under the hood: Kimi K2.5
Technically, Composer 2.5 shares the same base as Composer 2: Moonshot's open-source Kimi K2.5 checkpoint—a 1-trillion-parameter MoE model with about 32B activated per inference.
Cursor didn't train a foundation model from scratch; instead, it focused engineering on RL tuning. Three training improvements drove the gap:
- Text-feedback RL: Immediate hints on failed tool calls, not just end-of-sequence rewards, speeding up error correction.
- 25× synthetic tasks: Including "feature deletion" tasks that deliberately remove functionality for the model to restore.
- MoE infrastructure: Sharded Muon optimizer and dual-mesh HSDP for training efficiency.
Overall, 85% of compute budget went to RL and post-training, not base model training—a typical open-source LLM approach.
Next bet on xAI
Cursor's blog also dropped a bombshell: the next model is already in training, from scratch, no longer using Kimi as a base.
The partner is SpaceXAI (xAI's compute arm), using the Colossus 2 cluster—roughly 1 million H100-equivalent GPUs, 10× the compute of Composer 2.5.
Rumors had circulated since April that SpaceX acquired a $6 billion option to buy Cursor. Now the compute side is reported, signaling deep commitment from Musk and Cursor to the "AI coding" bet.
The real signal
Beyond scores and pricing, the most notable shift is Cursor's posture. Over the past two years, Cursor was often labeled a "wrapper"—a pretty IDE skin using Claude and GPT. The Composer series has started to tear off that label: first Composer 2 replaced most Claude calls for enterprise users, and now Composer 2.5 publicly benchmarks against Opus 4.7 at a fraction of the cost.
More critically, in the coding vertical, per-task cost dropping from over $10 to under $1 means agentic coding moves from "try it" to "let AI run all day" budget territory. A company that once dared only a dozen AI tasks per day can now run hundreds. That's the real workflow change.
As for the Terminal-Bench gap—Cursor will likely close it with the next model on SpaceX's Colossus 2. By the next release, the "wrapper" label may be gone for good.
Sources: CocoLoop, Cursor's Composer 2.5 matches Opus 4.7 and GPT-5.5 benchmarks at a fraction of the cost (The Decoder), Cursor released Composer 2.5 with up to 10x cost efficiency (Testing Catalog), Cursor Releases Composer 2.5, Matches Opus 4.7 On Some Benchmarks (OfficeChai)