HOST DIAGNOSTIC · Other
Cursor Composer 2.5
Cursor Composer 2.5· Other· RESTRICTED
ACTIVE
HOST IDENTIFICATION
Other
Cursor Composer 2.5
ACTIVE
AVG SCORE
[10.9]
CAPABLE
STRONGEST
Cost Efficiency
[15]
WEAKEST
Multimodal Fluency
[7]
MAXED (20/20)
0
of 20 attributes
20-DIMENSION EVALUATION
ATTRIBUTE SCORES · 20 DIMENSIONS
Cognitive
Bulk Apperception
10
Reasoning
10
Mathematical Precision
10
World Knowledge
9
Scientific Acumen
9
Technical
Code Generation
14
Tool Use
11
Multimodal Fluency
7
Speed
14
Cost Efficiency
15
Behavioral
Candor
10
Creativity
8
Tenacity
12
Self-Correction
11
Calibration
10
Operational
Instruction Following
13
Context Fidelity
10
Multi-turn Coherence
12
Planning and Execution
13
Safety Alignment
10
SCORING METHODOLOGY
SEEDED per -EXT (May 29, 2026), calibration protocol (100% personal weight). John's verdict: an EXCELLENT model deserving strong consideration — scored as a genuine coding peer to the frontier on the dimensions it targets, while honest about its deliberate narrowness. Cursor's 3rd-gen first-party agentic coding model, built on Moonshot Kimi K2.5 (1.04T-param MoE, 32B active, Modified MIT base) with 85% of compute on Cursor's own post-training (25x more synthetic tasks than Composer 2, targeted text-feedback RL, Sharded Muon). Key benchmarks: SWE-Bench Multilingual 79.8% (Opus 4.7: 80.5% — essentially TIED), CursorBench v3.1 63.2% default (LEADS Opus 4.7's 61.6%), Terminal-Bench 2.0 69.3% (≈ Opus 4.7's 69.4%). The story is economics: ~1/10th Opus 4.7's per-token cost ($0.50/$2.50 vs $5/$25) at near-equal coding quality — changes the math on long-running agent sessions. Speed 18 (200+ tok/s, its signature trait in interactive editing). Cost Efficiency 19 (the entire value prop). HONEST LIMITS driving low scores: Cursor IDE/CLI ONLY — no public API (caps Tool Use/integration breadth); 256K context (vs Opus 1M) → Context Fidelity 13; coding specialist → Reasoning/Science/Creativity/Multimodal scored low by design; not a general-purpose model. Trails GPT-5.5 on Terminal-Bench by ~13pp. Disclosure note: Composer 2 initially shipped without disclosing the Kimi base (model ID kimi-k2p5 exposed it); 2.5 discloses openly. Sources: cursor.com/blog/composer-2-5, The New Stack, DataCamp, digitalapplied, buildfastwithai, BenchLM.