GPT-5.6 Terra

GPT-5.6 Terra is a reasoning model from OpenAI in the GPT-5.6 family. 51 benchmarks count toward its score, in 8 categories.

availableShows if the model has enough results for an index.
IndexOverall score. 50 is the middle.73.4 ±2.3
CoverageShare of the index weight with results.100%
SpeedOutput tokens per second.36/s
Input / 1MUS dollars per 1M input tokens.$2 batch $1
Output / 1MUS dollars per 1M output tokens.$12 batch $6 US dollars per 1M output tokens in a batch.
ContextMaximum tokens in one request.1.05M
EloLMArena rating and rank.1446 (#52)

50 is the middle of the board. The range shows the doubt in the index. Batch work costs less.

28,119 votes. Elo shows what people prefer. It does not change the score.

CapabilitiesScore per category. 50 is the middle.

50 is the middle
AgenticMulti-step tasks with tools.
71.9
CodingCode writing and repair.
72.3
ReasoningLogic problems and puzzles.
79.7
MultimodalTasks with images and text.
64.5
KnowledgeFacts and expert knowledge.
69.1
MultilingualTasks in many languages.
83.7
InstructionTasks with strict rules in the prompt.
61.3
MathMath problems.
77.1

Results

51 counted
BenchmarkThe test name.CategoryThe capability that the test measures.ResultThe score from the publisher.IndexThis result on the index scale.RunThe settings of the run.DateDate of the result.Published byThe source of the result.
OTIS Mock AIME 2024-2025Math99.7%67.3max effortEpoch AI
ARC-AGI-1 (semi-private)Reasoning96.5%72.8max effortARC Prize Foundation
SWE-benchCoding95.4%74.11 Sept 2026Vals AI
GPQA diamondKnowledge93.3%64.4max effortEpoch AI
Graduate-Level Google-Proof Q&AKnowledge92.9%64.0David Rein et al.
GPQA DiamondKnowledge92.9%64.0David Rein et al.
Artificial Analysis GPQA DiamondKnowledge92.5%63.3Artificial Analysis
GPQA DiamondKnowledge90.9%62.21 Sept 2026Vals AI
IOICoding87.6%83.821 Sept 2026Vals AI
BrowseCompAgentic87.5%73.2OpenAI
VulcanBench v3Coding87.0%72.7VulcanBench contributors
MMLU ProKnowledge86.7%57.11 Sept 2026Vals AI
MMMU ProMultimodal86.5%68.11 Sept 2026Vals AI
τ²-Bench Tool-Agent-User EvaluationAgentic86.3%60.7Victor Barres et al.
FrontierMath-Tiers-1-3-v2-PrivateMath86.0%79.8max effortEpoch AI
LiveCodeBenchCoding85.9%62.41 Sept 2026Vals AI
ARC-AGI-2 (semi-private)Reasoning83.9%82.0max effortARC Prize Foundation
Artificial Analysis Long Context ReasoningReasoning83.0%65.7Artificial Analysis
MMMU-Pro with PythonMultimodal82.0%OpenAI
CyberGymAgentic81.8%72.1Zhun Wang et al.
LABBench2: An Improved Benchmark for AI Systems Performing Biology ResearchKnowledge81.2%Jon M. Laurent et al.
Massive Multi-discipline Multimodal Understanding ProMultimodal80.7%58.7MMMU-Pro authors
Artificial Analysis MMMU-ProMultimodal80.7%65.7Artificial Analysis
Terminal-Bench 2.1Agentic77.5%69.821 Sept 2026Vals AI
Artificial Analysis Coding IndexCoding76.7%73.0Artificial Analysis
Vibe Code Bench v1.1Coding74.6%73.2OpenHands21 Sept 2026Vals AI
ProofBench v1.1Math74.0%79.421 Sept 2026Vals AI
Artificial Analysis IFBenchInstruction71.2%61.3Artificial Analysis
FrontierMath-Tier-4-v2-PrivateMath70.7%81.9max effortEpoch AI
EuroEval SwedishMultilingual70.6%90.4EuroEval
DeepSWEAgentic69.6%74.4Datacurve AI
EuroEval FrenchMultilingual68.9%88.3EuroEval
EuroEval ItalianMultilingual67.8%87.0EuroEval
EuroEval PortugueseMultilingual65.4%83.9EuroEval
IOI v1Coding65.3%77.09 Aug 2026Vals AI
EuroEval DutchMultilingual65.0%83.4EuroEval
cursorBench32Coding64.9%73.1Benchmark authors
SWE-bench ProCoding63.4%65.3Xiang Deng et al.
EuroEval SpanishMultilingual62.1%79.8EuroEval
EuroEval GermanMultilingual59.7%76.9EuroEval
EuroEval PolishMultilingual59.0%75.9EuroEval
SkillsBenchCoding58.9%72.9OpenHands11 Sept 2026Vals AI
HealthBench ProfessionalKnowledge57.7%Rebecca Soskin Hicks et al.
FrontierCode 1.1 ExtendedCoding55.8%Cognition
Artificial Analysis SciCodeCoding55.0%69.2Artificial Analysis
Artificial Analysis Intelligence IndexKnowledge55.0%91.1Artificial Analysis
Furniture AssemblyReasoning54.2%78.1max effortEpoch AI
Chess PuzzlesReasoning54.0%93.3max effortEpoch AI
ToolathlonAgentic53.1%67.5OpenAI
HLE-VerifiedKnowledge51.1%Weiqi Zhai et al.
Artificial Analysis ITBench-AAAgentic51.0%Artificial Analysis
OSWorld 2.0Agentic50.2%79.8Mengqi Yuan et al.
Code MigrationCoding47.8%75.521 Sept 2026Vals AI
Artificial Analysis Omniscience AccuracyKnowledge46.8%71.8Artificial Analysis
GDPval-AA normalizedAgentic46.6%71.4Artificial Analysis
Artificial Analysis Agentic IndexAgentic43.7%72.8Artificial Analysis
SimpleQA VerifiedKnowledge43.2%61.5max effortEpoch AI
Artificial Analysis Humanity's Last ExamKnowledge42.9%71.1Artificial Analysis
APEX-Agents-AAAgentic38.9%71.8Artificial Analysis / Mercor
Mystery Game PuzzlesReasoning35.0%71.1max effortEpoch AI
HealthBench HardKnowledge32.7%73.2Meta AI
Critical Physics TasksReasoning30.0%95.0Artificial Analysis
Terminal-Bench 4.0Agentic26.3%78.021 Sept 2026Vals AI
ExploitGymAgentic23.2%77.8Zhun Wang et al.
Terminal-Bench 4.0.0Agentic21.5%74.6max effort · Codex21 Sept 2026Terminal-Bench
Terminal-Bench 3.0Agentic20.8%70.6Ryan Marten et al.
ApprenticeBench: end-to-end computer use, continual learning, and long-horizon agency on a real accounts-payable jobAgentic16.0%69.0NeoCognition
Vibe Code Bench 1-100Coding14.8%68.1OpenHands16 Sept 2026Vals AI
Agent Arena steerabilityAgentic4.972.9xhigh effort15 Sept 2026LMArena
Agent Arena command recoveryAgentic3.070.8xhigh effort15 Sept 2026LMArena
ARC-AGI-3 (semi-private)Reasoning0.8%max effortARC Prize Foundation
ProgramBenchCoding0.5%21 Sept 2026Vals AI
Agent Arena task outcomeAgentic-4.462.4xhigh effort15 Sept 2026LMArena

51 benchmarks count, from 65 of 73 results. A grey row does not count. Too few models took that benchmark.

Sources

BenchLM benchmark aggregationUsed with attribution; per-benchmark results credited to their original authorsOpenRouter, collected directlyNo licence statedEpoch AI, collected directlyCC BY — free to use and redistribute with attributionARC Prize Foundation, collected directlyNo licence stated. Their terms ask for written permission before commercial useVals AI, collected directlyNo licence stated. Read from the public leaderboard and credited to Vals AIEuroEval, collected directlyMIT — the leaderboard site and its CSV routes are in the licensed repositoryTerminal-Bench, collected directlyNo licence stated for the leaderboard. The harness repo is Apache-2.0LMArena, collected directlyCC BY 4.0 (lmarena-ai/leaderboard-dataset on Hugging Face)

Same level, lower price

DeepSeek V4.1 Flash70.6 · $0.6MiMo-V2.6-Pro73.5 · $0.87Step 5 Preview72.5 · $2.7

More from OpenAI

GPT-6 Astra83.5GPT-5.6 Sol76.7GPT-6 Sol76.9GPT-5.572.3GPT-5.5 Pro77.9