GPT-5.6 Luna

+ contribute

GPT-5.6 Luna is a closed-weights language model from OpenAI, released July 2026.

76.8±0.5#29 of 719 LLMsfrom 37 benchmark results across 10 sources

$ how it compares

The LLMs ranked either side of GPT-5.6 Luna, out of 828 with a score, then more from OpenAI.

#	model	score	where	RAM	released
27	MiMo-V2.6-Flash (309B-A15B) Xiaomi	77±1	api $0.17/M +local	~136GB	2026-09
28	GPT-6 Luna OpenAI	77±1	api $0.20/M		2026-09
29	GPT-5.6 Luna OpenAI	77±1	api $0.45/M		2026-07
30	Grok 4.6 SpaceXAI (xAI)	77±1	api $3.00/M		2026-08
31	DeepSeek V4 Pro (0813) DeepSeek	76±1	api $1.63/M +local	~1010GB	2026-08
	more from OpenAI
1	GPT-6 Astra OpenAI	87±1	api $20.00/M		2026-09
3	GPT-6.1 Sol OpenAI	85±1	api $4.00/M		2026-09
7	GPT-6 Sol OpenAI	83±1	api $4.00/M		2026-09

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #37 of 300 · 21% of its score · 9 tests

Trails GLM-5.3-Flash, Grok 4.7, Claude Sonnet 5 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena Agent (October 2026 aggregate IPS)	-0.0117 IPS ✎	▁▁▁▄▂▃▇█▂▂▁▃ beats 16/49	avg -0.312 IPS	▇▇ 5.1%	sharp
Terminal-Bench 4.0	17.3% ✎	█▅▅▅▅·▅▃▃▅▅▅ beats 5/20	very hard · avg 0.4%	▇ 2.9%	sharp
BrowseComp lab	83.3% ✎	▄··█▄··▄·▄█▄ beats 2/8	hard · avg 32%	▇ 2.7%	typical
Terminal-Bench 2.1 lab	84.7% ✎	▁·▁·▁▁▄▃▄█▃▇ beats 46/56	hard · avg 21%	▇ 2.6%	typical
SWE-bench Pro (split unspecified) lab	62.7% ✎	▃·▁▁▁▂▂▃█▇▃▁ beats 43/55	very hard · avg 5.1%	▇ 2.5%	sharp · at its level
Toolathlon (Tool Decathlon) lab	53.4% ✎	▃··▃··▃··▃█▅ beats 5/8	very hard · avg 13%	▇ 1.8%	at its level
DeepSWE v1.1 (resolved) lab	67.2% ✎	█··▃▃····▃·▃ beats 2/6	hard · avg 21%	▇ 1.5%	at its level
Agents' Last Exam (metric/harness unspecified) lab	50.3% ✎	▂·▂·█▂···▂▄▄ beats 8/11	very hard · avg 11%	▇ 1.0%	at its level

1 more test

test	result	vs the others in its run	difficulty	counts	why
KlinikeBench strict pass@1 (333 interactive clinical encounters, Terminus-2)	7.5% ✎	▁▁·▁▁▁▁▅▃█▅█ beats 7/27	very hard · avg 2.4%	▇ 0.8%	typical

coding▁▁▁▃▆██▇▆▄▃▁ #44 of 458 · 10% of its score · 4 tests

Beats GPT-5.5, GPT-5.4 here, though they rank higher overall. Trails Mistral Large 4 (public preview), Qwen3.8 27B, GLM-5.2 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena WebDev (Code Arena)	1518.0 ✎	▁▂▄▃▅█▄▇▆▄▁▁ beats 76/119	avg 981	▇▇ 4.5%	sharp
SWE-bench Pro (split unspecified) lab	62.7% ✎	▃·▁▁▁▂▂▃█▇▃▁ beats 43/55	very hard · avg 5.1%	▇ 2.5%	sharp · at its level
Luau Leetcode fully solved (Torpedo Software)	72.6% ✎	▁▁▁▁·▁▂▃▅▃█▇ beats 24/49	hard · avg 18%	▇ 1.6%	typical
DeepSWE v1.1 (resolved) lab	67.2% ✎	█··▃▃····▃·▃ beats 2/6	hard · avg 21%	▇ 1.5%	at its level

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #53 of 393 * · 2.6% of its score · 1 test

Beats MiMo-V2.6-Flash (309B-A15B), GPT-6 Luna here, though they rank higher overall. Trails Gemma 4 31B, Claude Sonnet 4.5, Gemini 3.5 Flash-Lite here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LMArena (text)	1453.1 ✎	▁▁▁▂▃▃▃▆▆▇█▃ beats 318/379	avg 1268	▇ 2.6%	blunt

math▁▂▃▅▆▅▄█▄▃▁▁ #15 of 398 · 40% of its score · 3 tests

Beats Qwen3.8 Max (2.4T-A95B), Claude Opus 4.8, GPT-5.4 here, though they rank higher overall.

test	result	vs the others in its run	difficulty	counts	why
FrontierMath tier 4 v2 +1	61.0% ✎	▁▂▁▂█▅▃▂▁▁▁▁ beats 51/63	very hard · avg 0.1%	▇▇▇▇▇▇▇▇▇ 23%	sharp · at its level · several runs
FrontierMath tiers 1–3 v2 +1	82.1% ✎	▂··▁▁▃▄▄█▅▃▃ beats 69/82	very hard · avg 1.2%	▇▇▇▇▇▇▇ 17%	sharp · at its level · several runs
OTIS mock AIME 2024–2025	98.3% ✎	▃▃▄▃▃▃▇▇█▃▃▅ beats 175/193	very hard · avg 8.5%	▇ 0.4%	sharp · maxed out · noisy

vision & documents▁▁▃▂▁▃▄▂▆▇█▄ #38 of 147 · 5.5% of its score · 5 tests

Beats MiMo-V2.6-Pro, MiMo-V2.6-Flash (309B-A15B) here, though they rank higher overall. Trails Qwen3.6 Flash, Qwen3.6 Plus, Qwen3.5 397B-A17B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena vision (style controlled)	1259.8 ✎	▁▁▂▂▂▂▄▅▅▆██ beats 109/145	avg 1068	▇ 1.8%	blunt
MMMU Pro lab	78.4% ✎	▅·····▃·▃·█▃ beats 3/7	hard · avg 40%	▇ 1.7%	typical
Arena document (style controlled)	1462.0 ✎	▁▁▃▃▄▃▇█▅▇▄▅ beats 19/37	avg 1335	▇ 0.9%	blunt
SciDocBench scientific-document workflows (overall, 496 instances)	50.6% ✎	▃▃···█▃▅▃·▃▅ beats 4/11	hard · avg 26%	▇ 0.8%	blunt · at its level
SpaceCast-Bench predictive spatial reasoning (overall, 16 task types)	46.6% ✎	▃·▅▃▅▃··▅··█ beats 7/11	very hard · avg 32%	▇ 0.3%	blunt · at its level · noisy

web search▂▁▂▆▄█▃▃▁▁·▁ #21 of 80 * · 2.7% of its score · 1 test

Beats GPT-5.5, Claude Opus 4.8, GPT-5.4 here, though they rank higher overall. Trails GPT-4o mini (2024-07-18), Mistral Small 3.1 24B (2503), GPT-4o (2024-08-06) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
BrowseComp lab	83.3% ✎	▄··█▄··▄·▄█▄ beats 2/8	hard · avg 32%	▇ 2.7%	typical

other tests15 not in a task · 24% of its score

test	result	vs the others in its run	difficulty	counts	why
Furniture Assembly (Epoch)	42.5% ✎	█▆▆▆▆▂▅▂▂▂·▅ beats 17/30	very hard · avg 1.2%	▇▇ 4.5%	sharp · at its level
Mystery Game Puzzles (Epoch)	21.0% ✎	█··▂▃▇▄█▂▃▁▁ beats 37/74	very hard · avg 9.4%	▇▇ 3.9%	sharp · at its level · noisy
KMMLU-HARD (Daekeun Kim, zero-shot)	75.8% ✎	█···▁▁▁▂▃·▆▂ beats 19/23	very hard · avg 29%	▇ 2.9%	sharp · at its level
Chess Puzzles (Epoch)	40.0% ✎	█▁▁▁▁▂▁▄▂▃▁▁ beats 120/135	very hard · avg 5.9%	▇ 2.4%	sharp · at its level · noisy
GPQA Diamond	91.6% ✎	▁▁▁▁▂▄▆▄▅▇█▅ beats 185/208	hard · avg 40%	▇ 1.6%	sharp
KoBALT-700 (Daekeun Kim)	77.0% ✎	▄▄█·····▄··▄ beats 3/5	hard · avg 37%	▇ 1.4%	typical
KMMLU-Pro (Daekeun Kim)	85.5% ✎	▄·▄█·····▄·▄ beats 2/5	medium · avg 40%	▇ 1.2%	typical
CLIcK Korean (Daekeun Kim)	91.5% ✎	▆▄▄▄█▄▂·▄█·▄ beats 19/23	medium · avg 58%	▇ 1.2%	typical

7 more tests

test	result	vs the others in its run	difficulty	counts	why
SimpleQA Verified	41.0% ✎	▁▁▃▂▃▃▆▆█▃▂▅ beats 35/78	very hard · avg 8.0%	▇ 1.0%	at its level
Icelandic LLM Leaderboard average (Miðeind, Sept 2026 suite)	53.9% ✎	▁▁▃▂▂▄▁▃▄▄▄█ beats 23/38	very hard · avg 14%	▇ 0.9%	at its level
HAE-RAE 1.0 (Daekeun Kim)	93.2% ✎	▄▄▄▂▂█▂▄▂█▃▄ beats 20/22	medium · avg 63%	▇ 0.9%	maxed out
MuSR Korean (Daekeun Kim)	71.5% ✎	█·····███·██ beats 0/5	medium · avg 42%	▇ 0.8%	typical
HealthBench Hard lab	31.4% ✎	▂··▂····▂▄·█ beats 4/8	very hard · avg 9.7%	▇ 0.6%	blunt · at its level
HealthBench lab	55.4% ✎	▃·▃·····█▃▃▅ beats 2/8	hard · avg 24%	▇ 0.6%	blunt · at its level
RobloxQA v2.0 multiple choice (Torpedo Software)	95.8% ✎	▁▁▃▁█▄▃▃▆▂▂▂ beats 32/49	easy · avg 78%	▇ 0.5%	maxed out

$ where to run it

Through an API

api · $0.45/M · $0.2 in / $1.2 out per 1M tokens · 1.1M context · platform.openai.com ↗ · price sourceStandard base token rate from models.dev. Provider-specific tiers, long-context surcharges, caching and promotions may differ.

$ how the score was worked out · Epoch AI — internal evaluatio… 36%, OpenAI vendor-reported benchm… 33%, LMArena leaderboard dataset 15% and 7 more

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

source	results	share of the score	licence
Epoch AI — internal evaluations	8	▇▇▇▇▇▇▇▇▇▇▇ 36%	CC BY 4.0
OpenAI vendor-reported benchmark scores for its own models (openai.com announcement posts and Deployment Safety Hub system cards) · lab-reported	11	▇▇▇▇▇▇▇▇▇▇ 33%	Factual vendor-reported score (not licensed content)
LMArena leaderboard dataset	5	▇▇▇▇ 15%	CC BY 4.0
Daekeun Kim Korean LLM benchmark runs	6	▇▇▇ 8%	MIT
Terminal-Bench 4.0 official leaderboard submissions	1	▇ 3%	Apache-2.0
Roblox LLM Leaderboard (Torpedo Software): RobloxQA v2.0 and Luau Leetcode	2	▇ 2%	MIT
Miðeind Icelandic LLM Leaderboard	1	▇ 1%	CC BY 4.0
SciDocBench: A Workflow-Centered Benchmark and Data Pipeline for Scientific Document Understanding (arXiv:2609.05141v2)	1	▇ 1%	CC BY 4.0
KlinikeBench: Evaluating Language Models Beyond Diagnostic Accuracy (arXiv:2609.38480v2)	1	▇ 1%	CC BY 4.0
SpaceCast-Bench: Evaluating Predictive Spatial Reasoning in Vision-Language Models (arXiv:2610.12402v1)	1	▇ 0%	CC BY 4.0
prior · pull toward the middle		▇ 0%
weighted average		= 76.8