GPT-5.4 (xhigh)

March 5, 2026Last synced Aug 1, 2026, 12:18 AM1050K context

Human Frontier

87.7

Human-calibrated frontier signal, backed by Arena-style preference evidence and separate from raw AgMoBench benchmark composite scores.

Blended Price

$5.63/M

Input Price

$2.50/M

Output Price

$15.00/M

Speed

—

TTFT

—

Benchmark Scores

Intelligence Index

Performance Radar

GPT-5.4 (xhigh)

Field Average

External Benchmarks

AA-Omniscience Accuracy

How GPT-5.4 (xhigh) Compares

Axes

X Axis

Y Axis

Bubble Size

Filters

Blended Price (USD)$0.00 – $37.5

AgMoBench Overall≥ 1.8

Providers

ai2ai21-labsalibabaanthropicarceeawsazurebaidu

Show quadrants

anthropic

openai

mistral

azure

cohere

meta

ai21-labs

inclusionai

google

ai2

ibm

alibaba

upstage

nvidia

aws

kimi

nous-research

thinking-machines

minimax

stepfun

tencent

deepseek

zai

multiversecomputing

xiaomi

xai

deepcogito

sarvam

kwaikat

baidu

swiss-ai-initiative

bytedance_seed

inception

reka-ai

arcee

perplexity

Bubble size = Context Window

8191.00

2000000.00

Compare with other models

GPQA Diamond

92.0 / 100

LiveCodeBench

— / 100

Terminal-Bench Hard

57.6 / 100

τ²-Bench

87.1 / 100

AA-Omniscience Indexaa_omniscience

5.7 / 100

CyberGymanthropic_launch

66.3 / 100

ARC-AGI-1arc_agi

93.7 / 100

ARC-AGI-1 Cost per Taskarc_agi

0.6

ARC-AGI-2arc_agi

74.0 / 100

ARC-AGI-2 Cost per Taskarc_agi

1.5

OSWorld (Unversioned Matrix Import)benchmark_matrix

75.0 / 100

SWE-bench Probenchmark_matrix

57.7

Terminal-Bench 2.0benchmark_matrix

75.1

AA-Omniscience Hallucination RatePredicted

93.3 / 100

Aider PolyglotPredicted

92.2 / 100

AIMEPredicted

0.3 / 30

AIME 2025Predicted

1.0 / 30

AlpacaEval 2.0 LCPredicted

29.9 / 100

AlpacaEval 2.0 RawPredicted

21.0 / 100

BFCL (Berkeley Function Calling)Predicted

59.3

BigCodeBench CompletePredicted

39.8 / 100

BigCodeBench InstructPredicted

31.2 / 100

AA Intelligence Index (Matrix)Predicted

60.4

AA Long Context Reasoning (Matrix)Predicted

76.3

AIME 2024Predicted

99.6

AIME 2025 (Matrix)Predicted

100.0

Arena-Hard AutoPredicted

85.2

BrowseCompPredicted

80.3

BRUMO 2025Predicted

99.7

CMIMC 2025Predicted

90.9

CritPtPredicted

23.4

GPQA Diamond (Matrix)Predicted

93.3

HLE (Matrix)Predicted

49.8

HMMT Feb 2025Predicted

97.7

HMMT Nov 2025Predicted

90.3

HumanEvalPredicted

96.2

IFBench (Matrix)Predicted

71.0

IFEvalPredicted

93.4

IMO 2025Predicted

45.4

LiveCodeBench (Matrix)Predicted

89.0

MATH-500 (Matrix)Predicted

98.9

MathArena Apex 2025Predicted

11.1

MMLU-Pro (Matrix)Predicted

88.7

MMMU-ProPredicted

80.9

MRCR v2Predicted

82.3

SimpleQAPredicted

69.4

SMT 2025Predicted

95.0

SWE-bench ProPredicted

52.8

Tau-Bench Telecom (Matrix)Predicted

98.8

Terminal-Bench 2.0Predicted

68.6

Terminal-Bench 1.0Predicted

53.8

USAMO 2025Predicted

39.4

Video-MMUPredicted

86.0

browsecompPredicted

80.3

Aider PolyglotPredicted

0.6

BIG-Bench HardPredicted

3.0

CAD-EvalPredicted

2.1

CyBenchPredicted

0.9

DeepResearchBenchPredicted

0.5

FictionLiveBenchPredicted

1.1

GeoBenchPredicted

0.0

GSM8K (Epoch)Predicted

2.4

HellaSwagPredicted

0.0

Lech Mazur WritingPredicted

8.8

METR Time HorizonsPredicted

51.9

The Agent CompanyPredicted

59.6

TriviaQAPredicted

25.6

WinoGrandePredicted

0.8

FrontierMathPredicted

39.4 / 100

GAIA Level 1Predicted

86.1

GAIA Level 2Predicted

74.4

GAIA Level 3Predicted

93.7

GAIAPredicted

73.9 / 100

LiveCodeBenchPredicted

0.9 / 100

LongBench v2 EasyPredicted

81.8

LongBench v2 HardPredicted

28.1

LongBench v2Predicted

24.7 / 100

MATH-500Predicted

0.9 / 100

MathVistaPredicted

73.3 / 100

MLE-bench (Original)Predicted

68.1 / 100

MMLU ProPredicted

0.9 / 100

MMMUPredicted

86.3 / 100

MMTU Table UnderstandingPredicted

70.0 / 100

MT-BenchPredicted

7.8 / 10

NoLiMa (NIAH)Predicted

93.7 / 100

OCRBench v2Predicted

87.0 / 100

Open LLM AveragePredicted

10.6 / 100

Open LLM: BBHPredicted

35.8 / 100

Open LLM: GPQAPredicted

26.2 / 100

Open LLM: IFEvalPredicted

22.6 / 100

Open LLM: MATH Level 5Predicted

1.3 / 100

Open LLM: MMLU-PROPredicted

21.1 / 100

Open LLM: MUSRPredicted

37.8 / 100

RE-BenchPredicted

100.0

SimpleBenchPredicted

72.5 / 100

simpleqaPredicted

64.9

SWE-bench LitePredicted

80.1 / 100

SWE-bench VerifiedPredicted

86.4 / 100

tau-bench RetailPredicted

91.5 / 100

WebArenaPredicted

0.2 / 100

WildBenchPredicted

36.4

BullshitBenchbullshitbench

48.0 / 100

Arena ELO: Codingchatbot_arena

1390.0

Arena ELO: Creative Writingchatbot_arena

1431.0

arena_elo_documentchatbot_arena

1470.0

Chatbot Arena ELOchatbot_arena

1466.0

arena_elo_visionchatbot_arena

1278.0

Apex Agentsepoch_ai

3.4

Arc Agi 2epoch_ai

0.7

Chess Puzzlesepoch_ai

0.4

Cl Benchepoch_ai

0.3

Cl Bench Lifeepoch_ai

0.2

DeepSWE (Epoch Import, Version Unverified)epoch_ai

70.5 / 100

Enigma Evalepoch_ai

0.2

Epoch Capabilities Indexepoch_ai

156.1

METR Time Horizonsepoch_ai

53.9

Mystery Game Puzzlesepoch_ai

0.4

OTIS Mock AIME 2024–2025epoch_ai

1.0

Posttrainbenchepoch_ai

0.0

Proofbenchepoch_ai

0.6

SimpleQA Verified (Epoch)epoch_ai

0.4

EQ-Bench 3eqbench

1328.0

FrontierMathfrontiermath

47.6 / 100

IDP Coreidp-leaderboard

84.4 / 100

IDP OlmOCRidp-leaderboard

81.0 / 100

IDP OmniDocidp-leaderboard

85.3 / 100

IDP Overallidp-leaderboard

83.5 / 100

LegalBenchlegalbench

86.0 / 100

LiveBench Codinglivebench

77.5 / 100

LiveBench Data Analysislivebench

79.3 / 100

LiveBench Instruction Followinglivebench

70.2 / 100

LiveBench Languagelivebench

82.6 / 100

LiveBench Mathlivebench

94.1 / 100

LiveBench Overalllivebench

80.3 / 100

LiveBench Reasoninglivebench

88.1 / 100

AIME 2026matharena

99.2

MedQA (USMLE)medqa

96.1

OSWorld (Unversioned Matrix Import)openai_launch

75.0 / 100

GDPvalopenai_launch

83.0 / 100

OpenRouter Token Volumeopenrouter_rankings

327000000000.0

OpenRouter Weekly Trendopenrouter_rankings

37.0

SWE-Bench Pro (Private)seal

43.4

SWE-Bench Pro (Public)seal

59.1

Vectara Factual Consistencyvectara_hallucination

93.0 / 100

Vectara Hallucination Ratevectara_hallucination

7.0 / 100

WeirdMLweirdml

57.4 / 100

ARC-AGI-3 Public Demoarc_agi

v3 · public-demo · GPT-5.4 (High) · effort high · 1 attempt · 5x-human-baseline-median-per-level · arc-agi-3:public-demo · record v3_Semi_Private:openai-gpt-5-4-2026-03-05-high

View run source

Source metadata

{
  "provider": "OpenAI",
  "modelType": "CoT",
  "modelsUrl": "https://arcprize.org/media/data/models.json",
  "modelGroup": "openai-gpt-5-4-2026-03-05-thinking",
  "modelNotes": null,
  "modelsEtag": "W/\"2fc19b6caba8b52765ca17eab04fc437\"",
  "scoreMetric": "relative-human-action-efficiency",
  "evaluationsUrl": "https://arcprize.org/media/data/evaluations.json",
  "evaluationsEtag": "W/\"9485c7ef206d9b12cdfdec6a631df56b\"",
  "upstreamModelId": "openai-gpt-5-4-2026-03-05-high",
  "environmentCount": 25,
  "modelReleaseDate": "2026-03-05T00:00:00.000Z",
  "upstreamDatasetId": "v3_Semi_Private",
  "actionBudgetPolicy": "5x-human-baseline-median-per-level",
  "technicalReportUrl": "https://arcprize.org/media/ARC_AGI_3_Technical_Report.pdf",
  "catalogMatchAliases": [],
  "agentHarnessPublished": false,
  "verificationAuthority": "ARC Prize Foundation",
  "upstreamRunIdPublished": false,
  "actualEvaluationCostUsd": 5187.41,
  "providerModelIdPublished": false
}

Evidence observed 2026-07-31

0.2 / 100

ARC-AGI-3 Public Demoarc_agi

v3 · public-demo · GPT-5.4 (High) · effort high · 1 attempt · 5x-human-baseline-median-per-level · arc-agi-3:public-demo · record v3_Semi_Private:openai-gpt-5-4-2026-03-05-high

View run source

Source metadata

{
  "provider": "OpenAI",
  "modelType": "CoT",
  "modelsUrl": "https://arcprize.org/media/data/models.json",
  "modelGroup": "openai-gpt-5-4-2026-03-05-thinking",
  "modelNotes": null,
  "modelsEtag": "W/\"2fc19b6caba8b52765ca17eab04fc437\"",
  "scoreMetric": "relative-human-action-efficiency",
  "evaluationsUrl": "https://arcprize.org/media/data/evaluations.json",
  "evaluationsEtag": "W/\"9485c7ef206d9b12cdfdec6a631df56b\"",
  "upstreamModelId": "openai-gpt-5-4-2026-03-05-high",
  "environmentCount": 25,
  "modelReleaseDate": "2026-03-05T00:00:00.000Z",
  "upstreamDatasetId": "v3_Semi_Private",
  "actionBudgetPolicy": "5x-human-baseline-median-per-level",
  "technicalReportUrl": "https://arcprize.org/media/ARC_AGI_3_Technical_Report.pdf",
  "catalogMatchAliases": [],
  "agentHarnessPublished": false,
  "verificationAuthority": "ARC Prize Foundation",
  "upstreamRunIdPublished": false,
  "actualEvaluationCostUsd": 5187.41,
  "providerModelIdPublished": false
}

Evidence observed 2026-07-31

0.2 / 100

ARC-AGI-3 Public Demoarc_agi

v3 · public-demo · GPT-5.4 (High) · effort high · 1 attempt · 5x-human-baseline-median-per-level · arc-agi-3:public-demo · record v3_Semi_Private:openai-gpt-5-4-2026-03-05-high

View run source

Source metadata

{
  "provider": "OpenAI",
  "modelType": "CoT",
  "modelsUrl": "https://arcprize.org/media/data/models.json",
  "modelGroup": "openai-gpt-5-4-2026-03-05-thinking",
  "modelNotes": null,
  "modelsEtag": "W/\"cfc9b87c645d3ef2103e359fd0f9f521\"",
  "scoreMetric": "relative-human-action-efficiency",
  "evaluationsUrl": "https://arcprize.org/media/data/evaluations.json",
  "evaluationsEtag": "W/\"0d65f6e0efe0d5cae15e8e2fe5a0994b\"",
  "upstreamModelId": "openai-gpt-5-4-2026-03-05-high",
  "environmentCount": 25,
  "modelReleaseDate": "2026-03-05T00:00:00.000Z",
  "upstreamDatasetId": "v3_Semi_Private",
  "actionBudgetPolicy": "5x-human-baseline-median-per-level",
  "technicalReportUrl": "https://arcprize.org/media/ARC_AGI_3_Technical_Report.pdf",
  "catalogMatchAliases": [],
  "agentHarnessPublished": false,
  "verificationAuthority": "ARC Prize Foundation",
  "upstreamRunIdPublished": false,
  "actualEvaluationCostUsd": 5187.41,
  "providerModelIdPublished": false
}

Evidence observed 2026-07-30

0.2 / 100

ARC-AGI-3 Public Demoarc_agi

v3 · public-demo · GPT-5.4 (High) · effort high · 1 attempt · 5x-human-baseline-median-per-level · arc-agi-3:public-demo · record v3_Semi_Private:openai-gpt-5-4-2026-03-05-high

View run source

Source metadata

{
  "provider": "OpenAI",
  "modelType": "CoT",
  "modelsUrl": "https://arcprize.org/media/data/models.json",
  "modelGroup": "openai-gpt-5-4-2026-03-05-thinking",
  "modelNotes": null,
  "modelsEtag": "W/\"bbdf5725ef7a703a7f0970e913262b6f\"",
  "scoreMetric": "relative-human-action-efficiency",
  "evaluationsUrl": "https://arcprize.org/media/data/evaluations.json",
  "evaluationsEtag": "W/\"6529094085f9ca7ce5f3c2e1957d3b9c\"",
  "upstreamModelId": "openai-gpt-5-4-2026-03-05-high",
  "environmentCount": 25,
  "modelReleaseDate": "2026-03-05T00:00:00.000Z",
  "upstreamDatasetId": "v3_Semi_Private",
  "actionBudgetPolicy": "5x-human-baseline-median-per-level",
  "technicalReportUrl": "https://arcprize.org/media/ARC_AGI_3_Technical_Report.pdf",
  "catalogMatchAliases": [],
  "agentHarnessPublished": false,
  "verificationAuthority": "ARC Prize Foundation",
  "upstreamRunIdPublished": false,
  "actualEvaluationCostUsd": 5187.41,
  "providerModelIdPublished": false
}

Evidence observed 2026-07-24

0.2 / 100

Senior SWE-Bench v2026.06senior_swe_bench

v2026.06 · official-100-task · GPT-5.4 (xhigh) · harness mini-swe-agent · effort xhigh · 1 attempt · pass@1 · harbor:snorkel-ai/senior-swe-bench-v2026.06 · provider openai/gpt-5.4 · record gpt-5-4-mini-swe · run c243aeb40c0c89716ec5a6e2056652f84b289975286619d6e25927d25cbb69c6

View run source

Source metadata

{
  "taskCount": 100,
  "agentGroup": "mini-swe-agent:gpt-5.4",
  "basicPassAt1": 41.05263157894737,
  "basicPassAt3": 54.736842105263165,
  "datasetPackage": "snorkel-ai/senior-swe-bench-v2026.06",
  "publicTaskCount": 50,
  "tastefulPassAt1": 16.842105263157894,
  "tastefulPassAt3": 27.368421052631582,
  "agentDescription": "OpenAI GPT-5.4 via mini-swe-agent harness (Portkey→OpenAI)",
  "privateTaskCount": 50,
  "upstreamProvider": "openai",
  "upstreamRunCount": 50,
  "averageAgentSteps": 86.26315789473684,
  "averageOutputTokens": 53854.221052631576,
  "averageTotalCostUsd": 3.071602650526316,
  "averageOutputCostUsd": 0.8078133157894736,
  "basicAllThreePassRate": 24.210526315789473,
  "tastefulAllThreePassRate": 6.315789473684211
}

Evidence observed 2026-07-22

16.8 / 100

Senior SWE-Bench v2026.06senior_swe_bench

v2026.06 · official-100-task · GPT-5.4 (xhigh) · harness mini-swe-agent · effort xhigh · 1 attempt · pass@1 · harbor:snorkel-ai/senior-swe-bench-v2026.06 · provider openai/gpt-5.4 · record gpt-5-4-mini-swe · run 90509a67fed6a6131978e82ddf4b66f2848116a98bd228326ce0482485a2feb3

View run source

Source metadata

{
  "taskCount": 100,
  "agentGroup": "mini-swe-agent:gpt-5.4",
  "basicPassAt1": 48.86363636363637,
  "datasetPackage": "snorkel-ai/senior-swe-bench-v2026.06",
  "publicTaskCount": 50,
  "tastefulPassAt1": 13.636363636363635,
  "agentDescription": "OpenAI GPT-5.4 via mini-swe-agent harness (Portkey→OpenAI)",
  "privateTaskCount": 50,
  "upstreamProvider": "openai",
  "upstreamRunCount": 50,
  "averageAgentSteps": 81.7840909090909,
  "averageOutputTokens": 50491.568181818184,
  "averageTotalCostUsd": 2.9705425000000005,
  "averageOutputCostUsd": 0.7573735227272728
}

Evidence observed 2026-06-30

13.6 / 100