AIEvals

Models

GPT-5.6 Luna

OpenAI’s smaller GPT-5.6 model for lower-cost reasoning and tool use.

OpenAIReleased 9 July 2026Context 1M tokensOutput 128k tokensPrivate weights$0.20 in / $1.20 out per Mtok

Family GPT-5.647 benchmarks with a published resultNo first place on the independent columns among the 46 models this site listsRead 17 October 2026

Where it stands

Rank compares tracked models with a score. Published shows the source’s rank and field size, when available.

BenchmarkScoreRankPublishedLeaderBehind
Vals IndexVals AI51.69%21of 3322of 41Gemini 4 Argon17.21 pts
Artificial Analysis Intelligence IndexArtificial Analysis3736of 42no list readClaude Opus 5.521 pts
Terminal-Bench 4.0Artificial Analysis11.62%33of 40no list readClaude Sonnet 5.552.02 pts
SWE-bench (Verified)Vals AI93.00%9of 239of 88Claude Opus 54.00 pts
GPQA DiamondVals AI91.67%16of 2318of 138Gemini 3.1 Pro3.78 pts
MMLU ProVals AI86.04%22of 2450of 138Claude Fable 5.16.34 pts
MMMU ProVals AI85.03%16of 2025of 93Claude Fable 5.15.61 pts
Vibe Code Bench v1.1Vals AI77.06%25of 3827of 106Claude Sonnet 5.515.33 pts
Terminal-Bench Science 0.1Artificial Analysis3.33%21of 24no list readGPT-6 Astra60.00 pts
ProgramBenchVals AI0.00%19of 2923of 54Claude Opus 5.518.50 pts
APEX-Agents-AAArtificial Analysis35.84%5of 7no list readGemini 3 5 Flash11.21 pts
τ³-BankingArtificial Analysis31.13%26of 30no list readQwen3.8 Max20.21 pts
ITBench-AAArtificial Analysis40.32%14of 18no list readGPT-5.6 Sol15.89 pts
MLCR-AAArtificial Analysis19.44%21of 37no list readClaude Sonnet 5.555.56 pts
ARC-AGI-3ARC Prize · Standard0.18%11of 11no list readGPT-6 Astra62.53 pts
Vibe Code Bench 1–100Vals AI22.59%5of 205of 20Claude Opus 5.57.77 pts
Harvey LAB-AAArtificial Analysis87.90%15of 20no list readMuse Spark 1.3 Max7.58 pts
EnterpriseOps-Gym-AAArtificial Analysis40.82%15of 22no list readClaude Fable 510.30 pts
AutomationBench-AAArtificial Analysis50.21%30of 40no list readGemini 4 Argon27.30 pts
SciCodeArtificial Analysis53.59%29of 40no list readClaude Opus 5.513.31 pts
GDP.pdfArtificial Analysis24.00%9of 40no list readGPT-6.1 Sol7.00 pts
CritPtArtificial Analysis20.57%19of 42no list readGPT-5.6 Sol11.72 pts
AA-OmniscienceArtificial Analysis-10.2840of 42no list readClaude Opus 5.556.7
AA-LCR v1.1Artificial Analysis83.67%9of 42no list readKimi K35.00 pts
Terminal-Bench 2.1Vals AI79.03%11of 3511of 75Claude Opus 5.58.61 pts
ProofBench v1.1Vals AI60.00%18of 3420of 46Claude Sonnet 5.540.00 pts
LegalBenchVals AI84.03%19of 2741of 149Claude Fable 54.53 pts
Harvey's Legal Agent BenchmarkVals AI1.25%34of 3749of 73Muse Spark 1.224.17 pts
SkillsBenchVals AI60.45%6of 207of 35DeepSeek V4.1 Flash9.35 pts
MysteryMechanismVals AI14.41%21of 2121of 21GPT-6 Astra38.74 pts
Vals Multimodal IndexVals AI69.06%7of 177of 33Claude Fable 55.09 pts
BioMysteryBenchVals AI61.48%17of 1918of 21Claude Sonnet 5.519.63 pts
Code MigrationVals AI44.55%19of 3719of 71Claude Sonnet 5.525.28 pts
CyberBench v1.1Vals AI73.63%8of 319of 43GPT-6 Sol4.35 pts
EMBVals AI67.12%16of 3716of 69Claude Fable 5.19.55 pts
Finance Agent (v2)Vals AI55.04%16of 3720of 73Gemini 4 Argon10.36 pts
IOIVals AI61.78%14of 3114of 38Gemini 4 Argon38.22 pts
Legal Research BenchVals AI36.54%30of 3734of 72Muse Spark 1.3 Max18.75 pts
MedCodeVals AI42.39%30of 3549of 104Claude Opus 521.18 pts
MedScribeVals AI84.39%26of 3634of 106Claude Opus 5.57.04 pts
Public Benefits Bench v1.1Vals AI61.16%23of 2927of 45Claude Opus 515.77 pts
SAGEVals AI44.22%25of 3245of 90Claude Opus 4.711.88 pts
Tax Agent BenchVals AI60.81%26of 3730of 64Claude Fable 5.116.83 pts
DeepSWE v1.1Datacurve67.19%10of 2010of 28GPT-6 Astra6.93 pts
GDPval-AA v2.1Artificial Analysis1,463.2327of 42no list readClaude Opus 5.5382.94
AA-Briefcase v1.1Artificial Analysis1,344.4927of 41no list readClaude Opus 5.5477.36
Humanity's Last ExamArtificial Analysis39.48%35of 42no list readClaude Opus 5.521.87 pts

Nearby on Vals Index

The two models above and below on the Vals Index (Vals AI), on every column where this model has a published number.

ModelVals Index%Artificial Analysis Intelligence IndexindexTerminal-Bench 4.0%SWE-bench%GPQA Diamond%MMLU Pro%MMMU Pro%Vibe Code Bench v1.1%Terminal-Bench Science 0.1%ProgramBench%
Grok 4.6xAI52.09%4417.17%95.60%94.70%89.40%–76.24%6.19%–
Claude Sonnet 5Anthropic51.77%3814.14%79.60%88.89%87.55%83.01%81.33%–0.00%
GPT-5.6 LunaOpenAI51.69%3711.62%93.00%91.67%86.04%85.03%77.06%3.33%0.00%
DeepSeek V4.1 FlashDeepSeek51.32%3926.77%––––84.74%9.05%–
Gemini 3.7 FlashGoogle51.27%3913.64%80.80%93.94%90.12%88.96%70.39%–0.00%
Bold marks a figure better than this model’s on the same column, whichever model it belongs to. Ranks cover the models shown from the stated source.

Cost and speed, as published

Additional measurements from Artificial Analysis’ own row. Read 30 September 2026.

As publishedCost per taskOutput speed
GPT-5.6 Luna (max)$0.18118.2 tok/s

Results by benchmark

Independent and vendor results retain their own harnesses and effort settings.

BenchmarkIndependent runAs reported by a publisher
APEX-Agents-AAKnowledge work35.84%Artificial Analysisnot published
τ³-BankingTool use31.13%Artificial Analysisnot published
ITBench-AAIT operations40.32%Artificial Analysisnot published
MLCR-AAMedical reasoning19.44%Artificial Analysisnot published
ARC-AGI-3Interactive reasoning0.18%ARC Prize · Standardnot published
Vibe Code Bench 1–100Application development22.59%Vals AInot published
Harvey LAB-AALegal reasoning87.90%Artificial Analysisnot published
EnterpriseOps-Gym-AAEnterprise workflows40.82%Artificial Analysisnot published
AutomationBench-AAWorkflow automation50.21%Artificial Analysisnot published
SciCodeScientific coding53.59%Artificial Analysisnot published
GDP.pdfDocument reasoning24.00%Artificial Analysisnot published
CritPtScientific reasoning20.57%Artificial Analysisnot published
AA-OmniscienceFactual knowledge-10.28Artificial Analysisnot published
AA-LCR v1.1Long-context reasoning83.67%Artificial Analysisnot published
Terminal-Bench 4.0Agentic coding11.62%Artificial Analysis
11.62%Vals AI
not published
Terminal-Bench 2.1Agentic coding79.03%Vals AI
80.90%Artificial Analysis
not published
SWE-bench (Verified)Software engineering93.00%Vals AInot published
GPQA DiamondScience reasoning91.67%Vals AI
91.11%Artificial Analysis
not published
MMLU ProKnowledge86.04%Vals AInot published
MMMU ProMultimodal reasoning85.03%Vals AI
78.55%Artificial Analysis
not published
Vibe Code Bench v1.1Product coding77.06%Vals AInot published
Terminal-Bench Science 0.1Agentic science3.33%Artificial Analysis
0.00%Vals AI
not published
ProgramBenchAgentic coding0.00%Vals AInot published
ProofBench v1.1Mathematics60.00%Vals AInot published
LegalBenchProfessional knowledge84.03%Vals AInot published
Harvey's Legal Agent BenchmarkProfessional knowledge1.25%Vals AInot published
SkillsBenchAgentic tooling60.45%Vals AInot published
MysteryMechanismScientific reasoning14.41%Vals AInot published
Vals Multimodal IndexComposite69.06%Vals AInot published
Vals IndexComposite51.69%Vals AInot published
BioMysteryBenchAgentic science61.48%Vals AInot published
Code MigrationAgentic coding44.55%Vals AInot published
CyberBench v1.1Security73.63%Vals AInot published
EMBFinancial modelling67.12%Vals AInot published
Finance Agent (v2)Financial analysis55.04%Vals AInot published
IOICompetitive programming61.78%Vals AInot published
Legal Research BenchProfessional knowledge36.54%Vals AInot published
MedCodeHealthcare administration42.39%Vals AInot published
MedScribeHealthcare administration84.39%Vals AInot published
Public Benefits Bench v1.1Public services61.16%Vals AInot published
SAGEEducation44.22%Vals AInot published
Tax Agent BenchProfessional knowledge60.81%Vals AInot published
DeepSWE v1.1Software engineering67.19%Datacurvenot published
GDPval-AA v2.1Knowledge work1,463.23Artificial Analysisnot published
AA-Briefcase v1.1Knowledge work1,344.49Artificial Analysisnot published
Humanity's Last ExamMultidisciplinary reasoning39.48%Artificial Analysisnot published
Artificial Analysis Intelligence IndexComposite37Artificial Analysisnot published
Independent column: the suite named under each figure, one harness per suite. Publisher column: each lab’s own table, reproduced on the benchmark’s page. Publisher ranks compare reported scores in the same table row only; ties share a rank and unreported cells are excluded.

Sources

Read on 17 October 2026.