HOST DIAGNOSTIC · OpenAI
GPT-5.6 Luna
GPT-5.6 Luna· GPT· PROPRIETARY
ACTIVE
HOST IDENTIFICATION
OpenAI
GPT-5.6 Luna
ACTIVE
AVG SCORE
[15.0]
STRONG
STRONGEST
Context Fidelity
[18]
WEAKEST
Candor
[8]
MAXED (20/20)
0
of 20 attributes
20-DIMENSION EVALUATION
ATTRIBUTE SCORES · 20 DIMENSIONS
Cognitive
Bulk Apperception
16
Reasoning
16
Mathematical Precision
17
World Knowledge
16
Scientific Acumen
16
Technical
Code Generation
15
Tool Use
16
Multimodal Fluency
16
Speed
12
Cost Efficiency
16
Behavioral
Candor
8
Creativity
13
Tenacity
17
Self-Correction
16
Calibration
9
Operational
Instruction Following
16
Context Fidelity
18
Multi-turn Coherence
15
Planning and Execution
18
Safety Alignment
14
SCORING METHODOLOGY
PROVISIONAL CALIBRATION per /NM310-A1 (2026-09-07) — 0 BAND / 3 BAND-COMPOSITE / 16 INHERITED-PROVISIONAL / 1 PROXY; re-calibrate when per-dimension evals land. Fast/low-cost tier of the GPT-5.6 family (preview 2026-06-26, GA 2026-07-09), replacing the old mini/nano suffix scheme. Pricing $1/$6 per M in/out (long-context $2/$9). Context ~1M, 128K max output. Model ID gpt-5.6-luna. Built for high-volume routine work: classification, extraction, summarizing, drafting, tagging. Not selectable in standard ChatGPT; available via Work, Codex + API. Agents' Last Exam ~50.3% (vendor-reported). Launch-cycle claim that 'Sol post-trained Luna' is hard to verify — see PostTrainBench benchmark record for the saturation-gap context. On Vibe Code Bench, ~4x more expensive and slower than Muse Spark 1.1 per AI Explained 2026-07-10. Parent Model wired to GPT-5.6 Sol as family anchor; Sol/Terra/Luna are capability-tier siblings, not fine-tunes. PROVENANCE (-A1 lanes M1/M1b, per-dim): Reasoning + Bulk Apperception [BAND-COMPOSITE:AA Intelligence Index — GPT-5.6 Luna (max) — 2026-07 51]; Code Generation [BAND-COMPOSITE:AA Coding Agent Index — GPT-5.6 Luna (max) — 2026-07 75]; Cost Efficiency [PROXY:AA $2.25/M blended]; all remaining 16 dimensions [INHERITED-PROVISIONAL:gpt-5-5]. Inheritance takes GPT-5.5, not the wired Parent Model GPT-5.6 Sol (Sol is a sibling tier, itself provisionally calibrated).