GPT-5.4

GPT-5.4 is a reasoning model from OpenAI. 62 benchmarks count toward its score, in 7 categories.

availableShows if the model has enough results for an index.
IndexOverall score out of 100.67.6 ±2.6
CoverageShare of the index weight with results.95%
SpeedOutput tokens per second.32/s
Input / 1MUS dollars per 1M input tokens.$2.5 batch $1.25
Output / 1MUS dollars per 1M output tokens.$15 batch $7.5 US dollars per 1M output tokens in a batch.
ContextMaximum tokens in one request.1.05M
EloLMArena rating and rank.1453 (#41)

The index is a score out of 100. The ± range shows how much it can change. Batch work costs less.

63,526 votes. Elo shows what people prefer. It does not change the score.

CapabilitiesScore per category, out of 100.

Out of 100
AgenticMulti-step tasks with tools.
63.9
CodingCode writing and repair.
66.1
ReasoningLogic problems and puzzles.
73.3
MultimodalTasks with images and text.
61.8
KnowledgeFacts and expert knowledge.
68.9
MultilingualTasks in many languages.
N/A
InstructionTasks with strict rules in the prompt.
67.5
MathMath problems.
70.3

Results

62 counted
BenchmarkThe test name.CategoryThe capability that the test measures.ResultThe score from the publisher.IndexThis result as a score out of 100.RunThe settings of the run.DateDate of the result.Published byThe source of the result.
τ²-Bench Tool-Agent-User EvaluationAgentic98.9%69.8Victor Barres et al.
AIMEMath96.7%58.816 Apr 2026Vals AI
OTIS Mock AIME 2024-2025Math95.3%64.8xhigh effortEpoch AI
LiveBench MathematicsMath94.1%73.7xhigh effort25 Jun 2026LiveBench
ARC-AGI-1 (semi-private)Reasoning93.7%71.4xhigh effortARC Prize Foundation
GPQA diamondKnowledge93.3%64.4xhigh effortEpoch AI
Graduate-Level Google-Proof Q&AKnowledge92.8%63.9David Rein et al.
GPQA DiamondKnowledge92.8%63.9David Rein et al.
Artificial Analysis GPQA DiamondKnowledge92.0%62.8Artificial Analysis
GPQA DiamondKnowledge91.7%62.91 Sept 2026Vals AI
LiveBench ReasoningReasoning88.1%78.4xhigh effort25 Jun 2026LiveBench
MMMU ProMultimodal87.5%69.81 Sept 2026Vals AI
LiveCodeBench ProCoding87.5%LiveCodeBench Pro authors
MMLU ProKnowledge87.5%58.41 Sept 2026Vals AI
ScreenSpot ProMultimodal85.4%67.5Kaixin Li et al.
React Native EvalsCoding85.3%66.8Callstack
LiveCodeBenchCoding84.1%60.81 Sept 2026Vals AI
CharXiv ReasoningMultimodal82.8%60.4CharXiv authors
BrowseCompAgentic82.7%69.2OpenAI
LiveBench LanguageKnowledge82.6%70.2xhigh effort25 Jun 2026LiveBench
MMMU-Pro with PythonMultimodal82.1%OpenAI
Artificial Analysis Long Context ReasoningReasoning82.0%65.0Artificial Analysis
Massive Multi-discipline Multimodal Understanding ProMultimodal81.2%59.5MMMU-Pro authors
LiveBench Data AnalysisReasoning79.3%66.1xhigh effort25 Jun 2026LiveBench
CyberGymAgentic79.0%70.2Zhun Wang et al.
FrontierMath-Tiers-1-3-v2-PrivateMath78.6%75.7xhigh effortEpoch AI
Artificial Analysis MMMU-ProMultimodal78.4%62.9Artificial Analysis
SWE-benchCoding78.2%60.31 Sept 2026Vals AI
LiveBench CodingCoding77.5%66.4xhigh effort25 Jun 2026LiveBench
MedXpertQA MultimodalMultimodal77.1%Meta AI
SWE-Bench verifiedCoding76.9%59.2high effortEpoch AI
OSWorld-VerifiedAgentic75.0%63.6Tianbao Xie et al.
ARC-AGI-2 (semi-private)Reasoning74.0%77.0xhigh effortARC Prize Foundation
Artificial Analysis IFBenchInstruction73.9%64.1Artificial Analysis
DeepSearchQAAgentic73.6%56.5Meta AI
Artificial Analysis Coding IndexCoding71.0%69.0Artificial Analysis
MCP AtlasAgentic70.6%61.3OpenAI
LiveBench Instruction FollowingInstruction70.2%70.8xhigh effort25 Jun 2026LiveBench
IOI v1Coding67.8%78.59 Aug 2026Vals AI
ERQAMultimodal65.4%60.0Qwen
Gert Labs Composite Game BenchmarkAgentic64.9%70.0Gert Labs
SimpleVQAMultimodal61.1%51.8Z.AI
Claw-EvalAgentic60.3%53.2Bowen Ye et al.
MedXpertQA TextKnowledge59.6%Meta AI
Terminal-Bench 2.0Agentic58.4%64.44 Jun 2026Vals AI
Vibe Code Bench v1.1Coding57.9%66.3OpenHands21 Sept 2026Vals AI
SWE-bench ProCoding57.7%59.8Xiang Deng et al.
ToolathlonAgentic54.6%68.9OpenAI
LiveBench Agentic CodingAgentic53.8%67.4xhigh effort25 Jun 2026LiveBench
OfficeQA ProMultimodal53.2%65.5OfficeQA Pro authors
Humanity's Last ExamKnowledge52.1%72.9Center for AI Safety et al.
SkillsBenchCoding51.7%66.8OpenHands11 Sept 2026Vals AI
Artificial Analysis Omniscience AccuracyKnowledge50.8%76.7Artificial Analysis
FrontierMath-Tier-4-v2-PrivateMath49.0%71.5xhigh effortEpoch AI
HealthBench ProfessionalKnowledge48.1%Rebecca Soskin Hicks et al.
FrontierMath-2025-02-28-PrivateMath47.6%75.0xhigh effortEpoch AI
SimpleQA VerifiedKnowledge45.1%63.3xhigh effortEpoch AI
Chess PuzzlesReasoning44.0%80.4xhigh effortEpoch AI
Artificial Analysis Humanity's Last ExamKnowledge43.7%71.9Artificial Analysis
ZeroBenchMultimodal41.0%Meta AI
HealthBench HardKnowledge40.1%82.8Meta AI
Humanity's Last Exam without toolsKnowledge39.8%62.5OpenAI
τ²-bench BankingAgentic39.4%27.1xhigh effort · Sierra4 Aug 2026Sierra Research
Artificial Analysis Intelligence IndexKnowledge39.0%71.1Artificial Analysis
JobBenchAgentic38.9%60.9Yuetai Li et al.
Furniture AssemblyReasoning37.5%66.2xhigh effortEpoch AI
Mystery Game PuzzlesReasoning37.0%73.3xhigh effortEpoch AI
GDPval-AA normalizedAgentic36.6%63.7Artificial Analysis
Code MigrationCoding35.0%67.321 Sept 2026Vals AI
APEX-Agents-AAAgentic33.3%67.4Artificial Analysis / Mercor
FrontierMath-Tier-4-2025-07-01-PrivateMath27.1%72.9xhigh effortEpoch AI
EBR-benchReasoning25.4%66.7xhigh effortEpoch AI
Critical Physics TasksReasoning23.4%87.4Artificial Analysis
MirrorCodeCoding15.6%high effortEpoch AI
ResearchClawBenchAgentic15.3%InternScience
ApprenticeBench: end-to-end computer use, continual learning, and long-horizon agency on a real accounts-payable jobAgentic11.0%66.2NeoCognition
ExploitGymAgentic6.0%65.6Zhun Wang et al.
Agent Arena command recoveryAgentic4.672.6high effort15 Sept 2026LMArena
Agent Arena steerabilityAgentic2.570.2high effort15 Sept 2026LMArena
ProgramBenchCoding0.5%21 Sept 2026Vals AI
ARC-AGI-3 (semi-private)Reasoning0.2%high effortARC Prize Foundation
ProgramBenchCoding0.0%21 Sept 2026Vals AI
Agent Arena task outcomeAgentic-1.665.5high effort15 Sept 2026LMArena

62 benchmarks count, from 72 of 83 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 statedVals AI, collected directlyNo licence stated. Read from the public leaderboard and credited to Vals AIEpoch AI, collected directlyCC BY — free to use and redistribute with attributionLiveBench, collected directlyNo licence stated for the leaderboard. The site repo serving these CSVs has no LICENSE; the harness repo carries upstream Apache-2.0 and MIT copies that cover code, not resultsARC Prize Foundation, collected directlyNo licence stated. Their terms ask for written permission before commercial useSierra Research, collected directlyMIT — results are in the licensed repositoryLMArena, collected directlyCC BY 4.0 (lmarena-ai/leaderboard-dataset on Hugging Face)

Same level, lower price

Qwen3.8-Flash-Next66.4 · Freedots3-note Preview65.5 · FreeGLM-5.3-Flash65.6 · $0.5

More from OpenAI

GPT-6 Astra83.5GPT-5.6 Sol76.7GPT-5.6 Terra73.4GPT-6 Sol76.9GPT-5.572.3