HOST DIAGNOSTIC · Other
SWE-1.6
SWE-1.6· Other· PROPRIETARY
ACTIVE
HOST IDENTIFICATION
Other
SWE-1.6
ACTIVE
AVG SCORE
[11.9]
CAPABLE
STRONGEST
Speed
[18]
WEAKEST
Multimodal Fluency
[6]
MAXED (20/20)
0
of 20 attributes
20-DIMENSION EVALUATION
ATTRIBUTE SCORES · 20 DIMENSIONS
Cognitive
Bulk Apperception
10
Reasoning
12
Mathematical Precision
10
World Knowledge
8
Scientific Acumen
9
Technical
Code Generation
14
Tool Use
15
Multimodal Fluency
6
Speed
18
Cost Efficiency
16
Behavioral
Candor
11
Creativity
8
Tenacity
13
Self-Correction
13
Calibration
11
Operational
Instruction Following
13
Context Fidelity
12
Multi-turn Coherence
13
Planning and Execution
14
Safety Alignment
11
SCORING METHODOLOGY
CALIBRATED. Cognition (same lab as FrontierCode benchmark, ), GA 2026-04-07 in Windsurf/Devin (Preview earlier; some sources cite a 2026-05-12 wider push). NARROW SOFTWARE-ENGINEERING AGENT, not a general chat model — post-trained from the SAME pre-trained base as SWE-1.5 via RL on real task environments using the Cascade agent harness on an open-source base. Independent/semi-independent sources: SWE-Bench Pro 50.4% (vs SWE-1.5's 40.1%; +~10pp, Cognition-reported, cross-checked vs Scale trajectories). Towards AI 18-task head-to-head corroborated UX gains. Speed up to 950 tok/s (Cerebras fast tier) / 200 t/s (Fireworks free tier) — industry-leading. DIMENSION RATIONALE: Code Gen 15 + Planning/Exec 15 + Tool Use 16 (parallel tool calls, less terminal reliance) + Tenacity 14 = focused SWE-agent strength. Speed 19 (950 t/s is top of the entire DB). Cost 17 (free tier for launch window). DELIBERATELY LOW generalist dims: World Knowledge 8, Scientific 9, Creativity 8, Multimodal 6 (Math 11) — this is a coding agent with no general-knowledge or multimodal remit; scoring it as a generalist would misrepresent it. Length-penalty training cut overthinking/looping/verbosity. No Artificial Analysis Index (not a general API model). NO sibling SWE-1.5 record in DB to anchor against; anchored to coding-specialist band + own benchmark position. Technologies relation LEFT EMPTY — no confirmed Cognition/Windsurf tech record; flagged for follow-up wiring.