Skip to content

OpenAI

GPT-6 Astra

OpenAI's strongest public model until GPT-6.2 Astra (17 November 2026). Scores 169.2 on the Epoch Capabilities Index. The first OpenAI model rated Critical for cybersecurity under its Preparedness Framework.

Announced
Access
Public

Benchmark scores

TestScoreBest on the trackerHow it was measured
Epoch Capabilities Indexmany tests combined into one score169 pointsGPT-6.2 Astra 174 pointsIndependentlisted as "GPT-6 Astra", CI 166.19 to 172.51Run 9 Nov 2026
AA Intelligence Index v5about ten tests in one score56.9 pointsClaude Fable 5.2 64.6 pointsIndependentv5.0.1; listed as GPT-6 Astra (Max)Run 10 Nov 2026
Remote Labor Indexpaid freelance projects21%Claude Fable 5.2 41%IndependentGPT 6 AstraRun 17 Sept 2026
FrontierSWE V2very long engineering jobs66%Claude Fable 5.2 70%Independentproximus harness, max effort, mean@5 over 34 tasks, 20-hour budget (best@5 73.0)
Terminal-Bench 4.1long command-line tasks60%Claude Opus 5.5 66%IndependentCodex harness, max effort, 330 trials, ±2.82 (95% CI)
MirrorCoderebuilding whole programs47%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, ±11.9 SERun 30 Aug 2026
APEX-Agentsprofessional office tasks65%Gemini 4 Carbon 84%Independentpass@1, unknown effort; listed as "GPT-6 Astra"
Humanity's Last Examexpert questions, no tools55%Top scoreIndependentleaderboard entry "GPT 6 Astra" (gpt-6-astra_unknown)
ARC-AGI-3puzzle games, standard harnessTypical first-time human player: 100%63%Claude Fable 5.2 81%IndependentARC Prize config "GPT-6 Astra (Max)" (openai-gpt-6-astra-max), standard harness
FrontierMath T1–3hard maths problems94%Claude Fable 5.2 95%Independentmax reasoning effort (gpt-6-astra_max)Run 30 Aug 2026
GPQA DiamondPhD-level science questionsPhD experts: 70%Saturated96%Gemini 4 Argon 96%Independentmax reasoning effort (gpt-6-astra_max)Run 30 Aug 2026
FrontierMath T4research-level maths problemsSaturated98%GPT-6.1 Sol 100%Independenthigh reasoning effort (gpt-6-astra_high)Run 30 Aug 2026
ARC-AGI-2visual puzzles, model aloneAverage person: 60%Saturated95%Top scoreIndependentARC Prize config "GPT-6 Astra (Max)" (openai-gpt-6-astra-max)
OTIS Mock AIMEcompetition mathsSaturated100%Top scoreIndependentmax reasoning effort (gpt-6-astra_max)Run 30 Aug 2026
METR time horizon (TH1.1 + TH2)task length an AI agent finishes half the time1.7 daysTop scoreIndependentPooled TH1.1 + TH2 suites; 95% CI 791.5–158701.0 min

Compare every model on the same tests →

Also on these charts

Sources