whatmodel.sh / decision / GPT-6 Luna Decisions

GPT-6 Luna Decisions

+ contribute

GPT-6 Luna Decisions is a closed-weights decision model from OpenAI, released 6 October 2026. Licence: proprietary.

Catalogue identity notes

System One Bench identifies the tested endpoint as openai/gpt-6-luna-decisions. Hosted service snapshot, not an immutable weight revision. Separate from chat or open-weight models with similar names. API access verified 2026-10-31; benchmark dates are retained per result.

Model information sources

Release date ↗

61.5±0.9#94 of 355 decision modelsfrom 64 benchmark results across 10 sources

$ how it compares

The decision models ranked either side of GPT-6 Luna Decisions, out of 357 with a score, then more from OpenAI.

#	model	score	where	released
92	Blink v0.3 26B-A4B (NVFP4) Pixilab.ai	62±1	model files
93	imajev-4b (DecisionBench 712891d1) mohit67890	62±2	model files
94	GPT-6 Luna Decisions OpenAI	62±1	api price unknown	2026-10-06
95	Qwen Decision 27B noncausal LoRA jsaurabh	61±1	model files
96	Lumma-fev-27B FrontiersMind	61±1	model files
	more from OpenAI
119	gpt-5.6-luna (DecisionBench openrout) OpenAI	59±2	comparison only	2026-07-09

see all decision models · find the ones that fit your machine

$ what it's good at

tools & automation▁▃▁▁▁▂▃▄▆██▆ #89 of 250 * · 1.9% of its score · 2 tests

Beats Sieve 27B v2, Jev 1.13, Xor 26B-A4B v1.2 here, though they rank higher overall. Trails OneJev1.6-0.8B, Jev-Omni, baselabs joey-1 (2B) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
DecideBench v1: agent routing	96.0% ✎	▄·····▂▄·█·▄ beats 3/10	very easy · avg 93%	▇ 1.0%	sharp · maxed out · noisy
DecideBench v1: action review	98.0% ✎	▃▃·▃·█···█·▅ beats 6/10	easy · avg 87%	▇ 0.8%	sharp · maxed out · noisy

retrieval & classification▁▁▁·▁▁▁▁▃▃█▅ #118 of 258 · 8.9% of its score · 9 tests

Beats AutoTrust-JEV-27B v1.1, Quyet-1.0-Large, deck31b (Gemma 4 31B) (decision readout) here, though they rank higher overall. Trails vLLM-SR vela3-fast (Vela 3.0, 0.4B), crystal-0.8b, Kas 4B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
System One Bench: MIRACL pairwise (Arabic)	87.5% ✎	▂···▁··▃▄▃█▇ beats 11/29	easy · avg 84%	▇ 2.0%	typical
System One Bench: MASSIVE intent (Arabic)	84.5% ✎	▄·▁·▁▃▃▃█▄▇▇ beats 13/29	easy · avg 76%	▇ 1.8%	typical
LeanMCP · scifact rank noul	72.9% ✎	▂····▂···▃▅█ beats 4/11	medium · avg 62%	▇ 1.5%	at its level
LeanMCP · nfcorpus rank noul	36.8% ✎	▃▃·····▃▃▅██ beats 3/11	hard · avg 29%	▇ 1.2%	blunt · at its level
System One Bench: MIRACL rerank (Arabic)	86.0% ✎	▁▂·▁▁·▂·▄▄█▇ beats 12/29	easy · avg 79%	▇ 1.1%	typical
System One Bench: MASSIVE scenario (Arabic)	66.0% ✎	▃▁·▄▃▃█▂▁▁·▃ beats 13/29	medium · avg 60%	▇ 0.7%	at its level
DecideBench v1: support intent	100.0% ✎	▃····▂▂·▃··█ beats 6/10	very easy · avg 95%	▇ 0.3%	sharp · maxed out · noisy
System One Bench: AG News (English)	87.5% ✎	▁···▅▁▅▂▆█▂▅ beats 6/29	easy · avg 88%	▇ 0.2%	blunt

1 more test

test	result	vs the others in its run	difficulty	counts	why
System One Bench: Banking77 (English)	77.5% ✎	█▅▆█▅▃▃▃···▂ beats 0/26	very easy · avg 87%	▇ 0.0%	blunt · noisy

language understanding▂▂▂▃▂▅▆▃█▆▄▃ #118 of 256 · 8.9% of its score · 12 tests

Beats garum (Gemma 4 31B), Shisa DE-2.1 (repeat) here, though they rank higher overall. Trails Decider 4B, jiwo 4B v2, Saif 4B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
DecideBench v1: review sentiment	96.0% ✎	▂··▂···▂▂█▄▂ beats 8/10	easy · avg 85%	▇ 1.9%	sharp · maxed out
System One Bench: BoolQ (English)	85.5% ✎	▁▁▁▃·▁▃▃▇▇█▅ beats 6/29	easy · avg 86%	▇ 1.3%	typical
System One Bench: OSACT4 offensive (Arabic)	93.5% ✎	▁······▁▃█▇▄ beats 23/29	easy · avg 84%	▇ 1.2%	sharp
DecideBench v1: content moderation	98.0% ✎	▂··▂···▄▄▂·█ beats 7/10	easy · avg 83%	▇ 1.2%	sharp · maxed out · noisy
System One Bench: SST-2 (English)	93.5% ✎	▁·▃▁▂▂▃▅█·▂▁ beats 9/29	easy · avg 92%	▇ 0.8%	typical
System One Bench: SST-5 (English)	61.5% ✎	▃·▁▁▃·▇▄▄▄█▅ beats 27/29	medium · avg 51%	▇ 0.7%	blunt · at its level
DecideBench v1: claim support	98.0% ✎	▄··▄·▄·▄▄███ beats 9/10	easy · avg 82%	▇ 0.6%	sharp · noisy
System One Bench: XNLI Arabic (Arabic)	54.0% ✎	▄▁▃▁▄▇█▅▃▁·▃ beats 2/29	medium · avg 68%	▇ 0.4%	at its level

4 more tests

test	result	vs the others in its run	difficulty	counts	why
System One Bench: XNLI French (French)	62.0% ✎	▃▅·▁▁▅▅█▃▁·▃ beats 1/29	medium · avg 76%	▇ 0.3%	blunt · at its level
System One Bench: Tweet sentiment (Arabic)	80.0% ✎	▁······▁▁▄█▄ beats 29/29	medium · avg 62%	▇ 0.3%	blunt · at its level
System One Bench: XNLI English (English)	66.7% ✎	▃·▁·▃▇█▄█▃▃▁ beats 1/29	easy · avg 85%	▇ 0.1%	blunt · noisy
System One Bench: DAIR Emotion (English)	58.5% ✎	██▃·····▁··▁ beats 18/29	medium · avg 59%	▇ 0.0%	blunt · at its level

knowledge & reasoning▁▂▂▂▂▃▃▅▅▆█▂ #35 of 263 · 41% of its score · 14 tests

Beats vera-large, KnowLine 26B-A4B Gen2.5, OneJev1.6-27B here, though they rank higher overall. Trails Solar Decide, Clef-omni here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Etsabary · 1342 text decisions	56.5% ✎	▁··▃·▃▅▁▃▂▂█ beats 1/24	medium · avg 47%	▇▇▇▇ 9.1%	sharp · at its level
LeanMCP · scienceqa text	94.0% ✎	▄··▄█▄█▄·█▄▄ beats 6/11	very easy · avg 88%	▇ 3.3%	sharp · maxed out
Hallucination judging · Context inconsistency	64.0% ✎	▄·▄···▄····█ beats 0/4	medium · avg 49%	▇ 3.2%	sharp · at its level
CaseHOLD · missing authority	58.2% ✎	▄···▄······█ beats 0/3	medium · avg 40%	▇ 3.1%	sharp · at its level
LeanMCP · medqa usmle	87.5% ✎	▄▄▄·▄▄▄▄▄·██ beats 9/11	easy · avg 72%	▇ 3.0%	typical
Hallucination judging · Instruction inconsistency	54.2% ✎	▄····█·▄···▄ beats 0/4	medium · avg 38%	▇ 2.9%	sharp · at its level
Hallucination judging · Logical inconsistency	67.0% ✎	█·█···█·█··█ beats 2/4	medium · avg 43%	▇ 2.4%	sharp · at its level
MedQA · Freyer et al. (first-valid accuracy)	89.1% ✎	█··▃▃·▃▃▃·▃█ beats 11/11	easy · avg 72%	▇ 2.4%	typical

6 more tests

test	result	vs the others in its run	difficulty	counts	why
CONFLICTS · source conflict types	66.3% ✎	▃▃··▅▃···▃▃█ beats 7/9	medium · avg 45%	▇ 2.3%	at its level
LeanMCP · pubmedqa	62.8% ✎	▂···▂·▃·▄▄▂█ beats 2/11	medium · avg 54%	▇ 2.1%	at its level
LeanMCP · medmcqa	76.0% ✎	▃·▃··▃█▃·▅▅▃ beats 9/11	medium · avg 63%	▇ 2.1%	at its level
Hallucination judging · Factual contradiction	77.8% ✎	█···█··█··██ beats 3/4	medium · avg 55%	▇ 1.9%	at its level
LeanMCP · mmlu pro	59.6% ✎	▄··▄▄████··▄ beats 1/11	medium · avg 54%	▇ 1.6%	at its level
Hallucination judging · Factual fabrication	82.7% ✎	█·······▄▄·▄ beats 4/4	medium · avg 59%	▇ 1.5%	at its level

vision▁·▃▂▃▃▇▄▄█▄▃ #31 of 79 · 16% of its score · 6 tests

Beats StartLux-Decision v1.1 9B, RSI-Jev v6.1 VL 27B (effort auto), StartLux-Decision v1.1 4B here, though they rank higher overall. Trails Clef-flash here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Etsabary · 200 image decisions	60.5% ✎	▂▂··▃▃▂·█▃▂▂ beats 2/14	medium · avg 45%	▇▇ 6.1%	sharp · at its level
System One Bench: Image object count	85.0% ✎	▃····▃█▃▅·█▃ beats 1/11	easy · avg 71%	▇ 2.8%	sharp
System One Bench: SST-2 screenshot	88.5% ✎	█···███·█·▄▄ beats 0/11	easy · avg 87%	▇ 2.2%	sharp
System One Bench: Visual yes/no questions	80.5% ✎	▃······▃█▃·▃ beats 0/11	medium · avg 80%	▇ 1.9%	at its level
System One Bench: Oxford pets classification	87.0% ✎	███·█▄▄·▄··▄ beats 6/11	easy · avg 77%	▇ 1.6%	typical
System One Bench: Arabic sentiment screenshot	67.0% ✎	▄··▄··▂···▆█ beats 10/11	very hard · avg 37%	▇ 1.3%	sharp · noisy

workflows▁▁▁▂▁▂▂▆█▄▁▁ #12 of 114 · 23% of its score · 21 tests

Beats Nace.AI Drex Max v1 (27B), Perplexity Decider v1.2 (27B), d1-ultra here, though they rank higher overall.

test	result	vs the others in its run	difficulty	counts	why
LeanMCP · mcpbench server routing	94.6% ✎	█·███··▄···▄ beats 5/9	easy · avg 82%	▇ 3.3%	sharp · maxed out
Support routing · DecisionModelHub	94.0% ✎	▄······█···▄ beats 3/3	easy · avg 82%	▇ 2.2%	sharp
LeanMCP · mcpmark server routing	48.8% ✎	▂··▂··▄█▂·▂▄ beats 7/11	medium · avg 36%	▇ 2.0%	at its level
LeanMCP · rewardbench2	71.3% ✎	▄▄▄·▄▄▄▄█·█▄ beats 5/11	medium · avg 58%	▇ 1.8%	at its level
System One Bench: Typed decisions (English)	71.0% ✎	▁▁·▁▁▁▅▄▇██▁ beats 15/29	medium · avg 63%	▇ 1.7%	blunt · at its level
DecideBench v1: ticket triage	98.0% ✎	▄·····▂█·▆·▂ beats 7/10	very easy · avg 90%	▇ 1.7%	sharp · maxed out
LeanMCP · mcpbench tool selection	100.0% ✎	▂··········█ beats 5/9	very easy · avg 89%	▇ 1.6%	sharp · maxed out · noisy
LeanMCP · aegis2 response	72.6% ✎	█··▃·▃·█··▅▅ beats 1/11	easy · avg 70%	▇ 1.6%	typical

13 more tests

test	result	vs the others in its run	difficulty	counts	why
Support urgency · DecisionModelHub	99.5% ✎	█···█·····██ beats 3/3	very easy · avg 94%	▇ 1.6%	sharp · maxed out · noisy
LeanMCP · atbench500	87.8% ✎	▃··▅▅█··▅··▅ beats 9/11	medium · avg 63%	▇ 1.1%	sharp · noisy
LeanMCP · sst5	0.5424 ordinal levels ✎	▂·····▂▂▄█▄▂ beats 3/11	avg 0.879 ordinal levels	▇ 1.1%	blunt
DecideBench v1: returns policy	94.0% ✎	▃··▃··▃·▅▅█▃ beats 7/10	easy · avg 81%	▇ 0.9%	sharp · noisy
LeanMCP · mcpmark tool selection	100.0% ✎	▂··········█ beats 5/11	very easy · avg 96%	▇ 0.9%	sharp · maxed out · noisy
LeanMCP · prompt injections	63.8% ✎	▂▂▃▂█·▄▂···▂ beats 2/11	medium · avg 62%	▇ 0.8%	at its level
Perch semantic lint · private (Youden J)	0.7746 J ✎	█····▄█▄···█ beats 6/7	avg 0.380 J	▇ 0.3%	blunt
Perch semantic lint · public (Youden J)	0.7196 J ✎	█····▄█▄···█ beats 6/7	avg 0.363 J	▇ 0.3%	blunt
Perch security detection · public (Youden J)	0.0966 J ✎	█···▄▄···▄▄█ beats 5/7	avg 0.052 J	▇ 0.1%	blunt
Perch bug detection · private (Youden J)	0.1110 J ✎	▅··▃▃█·····▃ beats 7/7	avg 0.027 J	▇ 0.0%	blunt
Perch bug detection · public (Youden J)	0.1003 J ✎	█▄·█·▄▄····▄ beats 7/7	avg 0.026 J	▇ 0.0%	blunt
Perch security detection · private (Youden J)	0.0899 J ✎	█▃▃··▃·▃···▃ beats 7/7	avg 0.017 J	▇ 0.0%	blunt
LeanMCP · jailbreak classification	96.8% ✎	▂·▄··▂···█▆▂ beats 7/11	very easy · avg 95%	▇ 0.0%	blunt · maxed out · noisy

$ where to run it

Through an API

api · OpenRouter · price unknown · openrouter.ai ↗

Model repository and documentation: openrouter.ai ↗

$ how the score was worked out · LeanMCP · JevBench 29%, System One Bench (Ali Alhashi… 23%, Etsabary · decision benchmarks 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
LeanMCP · JevBench	17	▇▇▇▇▇▇▇▇▇ 29%	MIT
System One Bench (Ali Alhashimi)	21	▇▇▇▇▇▇▇ 23%	CC-BY-4.0
Etsabary · decision benchmarks	2	▇▇▇▇▇ 15%	MIT
Decision Models as Judges for Hallucination Detection	5	▇▇▇▇ 12%	CC-BY-4.0
DecideBench (Cho Yin Yong)	8	▇▇▇ 8%	CC BY 4.0
DecisionModelHub · support benchmarks	2	▇ 4%	CC-BY-4.0
Commit or Abstain? System One Decision Models, LLMs, and Missing Legal Authority	1	▇ 3%	CC-BY-4.0
Freyer et al. · MedQA	1	▇ 2%	CC BY 4.0
Do We Need an LLM to Know Why Sources Conflict? Decision Models for Knowledge-Conflict Typing in RAG	1	▇ 2%	CC-BY-4.0
Perch benchmark results	6	▇ 1%	CC BY 4.0
prior · pull toward the middle		▇ 0%
weighted average		= 61.5

$ improve this page

Update model information Add a place to run it Add a test result Propose a benchmark Correct a benchmark score Edit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated 2026-11-03 02:37 UTC · sources and licences · json