Skip to content

OpenAI

GPT-5.4

Takes prompts of 1 million tokens and can operate a computer itself. Released with GPT-5.4 Pro.

Announced
Access
Public

Benchmark scores

TestScoreBest on the trackerHow it was measured
Epoch Capabilities Indexmany tests combined into one score158 pointsGPT-6.2 Astra 174 pointsIndependentlisted as "GPT-5.4", CI 155.95 to 159.96Run 9 Nov 2026
MirrorCoderebuilding whole programs16%Claude Opus 5.5 79%Independenthigh reasoning effort, MirrorCode (ML, +Private, 2L): 15 programs x 2 languages, 3 runs each, 10B-token / 7-day budget, ±7.7 SERun 10 Aug 2026
APEX-Agentsprofessional office tasks52%Gemini 4 Carbon 84%Independentpass@1, unknown effort; listed as "GPT-5.4"
Humanity's Last Examexpert questions, no tools36%GPT-6 Astra 55%Independentleaderboard entry "gpt-5.4-2026-03-05 (xhigh thinking)" (gpt-5.4-2026-03-05_xhigh)
ARC-AGI-3puzzle games, standard harnessTypical first-time human player: 100%0.2%Claude Fable 5.2 81%IndependentARC Prize config "GPT-5.4 (High)" (openai-gpt-5-4-2026-03-05-high), standard harness
FrontierMath T1–3hard maths problems79%Claude Fable 5.2 95%Independentxhigh reasoning effort (gpt-5.4-2026-03-05_xhigh)Run 11 Jun 2026
GPQA DiamondPhD-level science questionsPhD experts: 70%Saturated93%Gemini 4 Argon 96%Independentxhigh reasoning effort (gpt-5.4-2026-03-05_xhigh)Run 6 Mar 2026
FrontierMath T4research-level maths problemsSaturated49%GPT-6.1 Sol 100%Independentxhigh reasoning effort (gpt-5.4-2026-03-05_xhigh)Run 11 Jun 2026
ARC-AGI-2visual puzzles, model aloneAverage person: 60%Saturated74%GPT-6 Astra 95%IndependentARC Prize config "GPT-5.4 (XHigh)" (gpt-5-4-xhigh)
OTIS Mock AIMEcompetition mathsSaturated98%GPT-5.5 100%Independenthigh reasoning effort (gpt-5.4-2026-03-05_high)Run 15 Jul 2026
SWE-bench Verifiedfixing real software bugsKnown flaws77%Claude Opus 4.7 83%Independenthigh reasoning effort (gpt-5.4-2026-03-05_high)Run 6 Mar 2026
METR time horizon (TH1.1 + TH2)task length an AI agent finishes half the time5.9 hGPT-6 Astra 1.7 daysIndependentPooled TH1.1 + TH2 suites; 95% CI 199.5–592.0 min
METR time horizontask length an AI agent finishes half the timeNo longer updated5.7 hClaude Mythos Preview 17 hIndependentMETR Horizon v1.1 run 'gpt_5_4', xhigh reasoning effort (per Epoch AI mapping); 95% CI 186.6–768.8 min

Compare every model on the same tests →

Sources