OpenAImodel systemcard

GPT-5.6 Luna

#13of 27 · general

effort ladder collapsed

1.1Mcontext

Proprietarynon-commercial

2026-07-09released

2026-02-16knowledge cutoff

Standing

dot = this model · shaded = the field's middle 60% · track = full field

General

49.7#13/27

Math

41.6#15/27

Coding

39.0#15/27

Health

27.3#20/26

Engineering

31.2#12/17

Cyber

92.3#2/14

Legal

66.8#5/26

Instruct

3.1#25/27

Agents

64.2#8/27

Professional

42.2#15/27

Multimodal

38.6#12/19

Hallucination

19.3#21/27

Speed

97.4#1/22

Value

57.8#7/27

Measured

≈ imputed · amber dot = lab-self-reported among the sources

GPQA Diamond91.6%#13

HLE39.5%#17

MMLU-Pro86.0%#20

LiveBench73.6%#13

SWE-bench V93.0%#5

SciCode52.5%#15

CritPt20.6%#8

Terminal-Bench80.0%#7

APEX Agents35.8%#8

GDPval54.0%#8

AA Index52.3#11

Vals Index69.9#5

Epoch Capabilities Index156.1#7

Operations

$0.20 / $1.2$ per 1M tokens, in / out

$0.45blended

195tokens / sec

$0.17$ per solved task

~1477arena elo · predicted, never voted on

Series

OpenAI's models on the board, by release

GPT-5.52026-04-23 · 67GPT-5.6 Sol2026-07-09 · 85GPT-5.6 Terra2026-07-09 · 59GPT-5.6 Luna2026-07-09 · 50

Neighbors

nearest capability fingerprints · deltas describe the neighbor

GPT-5.6 Terra→ stronger math · pricier Grok 4.5→ follows instructions better · pricier HY3→ weaker agent · cheaper Claude Sonnet 5→ follows instructions better · pricier Kimi K2→ near-twin