HOST DIAGNOSTIC · Other
Cursor Composer 2
Cursor Composer 2· Other· RESTRICTED
ACTIVE
HOST IDENTIFICATION
Other
Cursor Composer 2
ACTIVE
AVG SCORE
[10.5]
CAPABLE
STRONGEST
Speed
[14]
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
10
Scientific Acumen
9
Technical
Code Generation
12
Tool Use
11
Multimodal Fluency
7
Speed
14
Cost Efficiency
14
Behavioral
Candor
10
Creativity
8
Tenacity
11
Self-Correction
11
Calibration
9
Operational
Instruction Following
12
Context Fidelity
10
Multi-turn Coherence
11
Planning and Execution
12
Safety Alignment
9
SCORING METHODOLOGY
CALIBRATED. Anysphere/Cursor's 3rd-gen first-party agentic coding model, released 2026-03-19. Fine-tuned from Moonshot Kimi K2.5 via Cursor's first continued-pretraining run + long-horizon RL (the Kimi base initially UNDISCLOSED — dev found 'kimi-k2p5' in model ID, prompting transparency debate; 2.5 later disclosed openly). Anchored DIRECTLY below Composer 2.5 (in-graph gold standard). Benchmarks (Cursor-published): CursorBench 61.3, Terminal-Bench 2.0 61.7% (beat Opus 4.6's 58.0%, trailed GPT-5.4's 75.1%), SWE-Bench Multilingual 73.7%. Pricing $0.50/$2.50 per M standard (~86% cheaper than Composer 1.5); Fast variant $1.50/$7.50 (default). 200K context, 200+ tok/s. DIMENSION RATIONALE: Code Gen 16 (one below 2.5's 17 — CursorBench 61.3 vs 63.2, near-peer); Speed 18 + Cost 18 (signature traits); deliberately LOW generalist dims (World 13/Sci 11/Creativity 10/Multimodal 9) — coding specialist, same honest-narrowness pattern as 2.5. Tool Use 14 (Cursor IDE/CLI only, no public API). Michael Spencer: 'great at coding... doesn't have the reasoning depth Opus 4.6 has.' Sources: cursor.com/blog/composer-2, VentureBeat, buildfastwithai, LLMReference (arxiv 2603.24477). Anchored against Cursor Composer 2.5.