Atlas

Models

OpenAI

GPT-6.2 Astra

170.1
Capability · 95% interval 168.0–172.3
Rank 3 of 899 (interval 2–3)
Relative cost 2.2× · $0.23 per short-answer task · $1.48 per agentic task

GPT-6.2 Astra is OpenAI's November 17, 2026 frontier reasoning model for complex professional work, coding, computer use and documents. It accepts text and image inputs and produces text, with low, medium, high, xhigh and max reasoning effort; the API rejects effort none. Tool calling uses the Responses API. Its documented total context window is 1,050,000 tokens, with at most 922,000 input tokens and 128,000 output tokens. List price is $8 input, $0.40 cached input and $40 output per million tokens, 20% below GPT-6 Astra.

Scored on 220 benchmarks · 414 total scores · 370 independently reported

Score providers

Organizations reporting headline scores for 220 benchmarks

GPT-6.2 Astra benchmark scores

In fitin the capability model (41 of 220)
Legal · 69 benchmarks · first place on 5
Legal Research Bench Health
1 / 88———
Harvey LAB-AA v1.1 (hallucination-gated)
1 / 28———
HLAB-Vals-1.1 - Bankruptcy Restructuring
1 / 40———
HLAB-Vals-1.1 - International Trade Sanctions
1 / 40———
HLAB-Vals-1.1 - Structured Finance Securitization
1 / 40———
Finance & Accounting · 35 benchmarks · first place on 1
Excel Modeling Benchmark v1.1 Dataroom
1 / 44———
Excel Modeling Benchmark v1.1 Comps
2 / 44———
Excel Modeling Benchmark v1.1 Scratch
3 / 44———
Excel Modeling Benchmark v1.1In fit
4 / 44~2 mo~Dec 2027—
Excel Modeling Benchmark v1.1 Operating
4 / 44———
Code · 18 benchmarks · first place on 6
Vibe Code Bench 1–100 v1.1In fit
1 / 8~1 mo~Mar 2028—
Terminal-Bench 4.0 — Software
1 / 59———
Terminal-Bench 4.0 — Hardware
1 / 59———
AA-3D v1.0 EloIn fitnot reported
1 / 27~1 mo—2026-11-17
OpenAI research debugging - internal release scorenot reported
1 / 8——2026-11-17
Agents · 15 benchmarks · first place on 4
Agents' Last Exam v1.1 Scorex-high
1 / 19——2026-11-17
AutomationBench 1.0.6 HR
1 / 2——2026-11-17
AutomationBench-AA Operations, HR, Marketing, Sales and Support
1 / 26———
LifeVal (GPT-6.2 launch chart)
1 / 4——2026-11-17
AutomationBench 1.0.6 (Zapier)In fit
2 / 42~2 mo~Sep 20272026-11-17
Indexes · 15 benchmarks · first place on 7
Vals Science Index
1 / 24———
Epoch Capabilities Index (ECI)not reported
1 / 255———
Artificial Analysis Intelligence Index v5.0.1
1 / 67———
Artificial Analysis Strategy & Ops Index
1 / 26———
Artificial Analysis Healthcare & Medical Index
1 / 26———
General QA · 12 benchmarks · first place on 8
GPQA DiamondIn fitnot reported
1 / 410~5 moApr 20262026-11-17
ARC-AGI-2In fit
1 / 102~3 moSep 2026—
ARC-AGI-3 (Semi-Private)In fitx-high
1 / 30~2 mo~Jan 2027—
ARC-AGI-1 Semi-PrivateIn fithigh
1 / 95~5 moFeb 2026—
AA-Omniscience v1.1 Accuracy
1 / 76———
Science · 12 benchmarks · first place on 7
Terminal-Bench-Science 0.1In fitx-high
1 / 78~1 mo~Mar 2027—
MysteryMechanismIn fit
1 / 31~2 mo~Jan 2028—
Terminal-Bench Science 0.1 (Vals)In fit
1 / 44~1 mo~Apr 2027—
Terminal-Bench Science — Physical Sciences (Vals)
1 / 44———
Terminal-Bench Science — Earth Sciences (Vals)
1 / 44———
Chat & Writing · 8 benchmarks
AA-PersonalityBench Conviction indexlow
11 / 43———
AA-PersonalityBench Warmth index
17 / 43———
AA-PersonalityBench Imagination index
20 / 43———
AA-PersonalityBench Expressiveness index
22 / 43———
AA-PersonalityBench Transparency index
29 / 43———
Cybersecurity · 8 benchmarks · first place on 5
ExploitBench - release-reported exploit progress
1 / 18——2026-11-17
DeepsecBench-AA (Artificial Analysis)medium
1 / 26———
ExploitBench - Internal Port (June-August 2026)
1 / 7——2026-11-17
SecureSWE (OpenAI internal)
1 / 4——2026-11-17
PatchBench (OpenAI internal)
1 / 4——2026-11-17
Professional Work · 8 benchmarks · first place on 2
AA-AnalystAgent (pass^5)In fit
1 / 53~2 mo~Dec 2027—
GDPval v2 (OpenAI internal)
1 / 7——2026-11-17
MLCR-AA Accuracy (October 2026 methodology)
3 / 114———
AA-AnalystAgent (pass@1)x-high
5 / 53———
GDP.pdf (Artificial Analysis, Mean Criteria)
5 / 162———
Games · 6 benchmarks · first place on 5
Mystery Game PuzzlesIn fit
1 / 82~4 mo~Nov 2026—
EBR-bench v4 (card ban, single agent)In fit
1 / 15~1 moSep 2026—
SpireClimbIn fit
1 / 10~1 mo~Sep 2027—
SpireClimb — Slay the Spire 1, Ascension 20 Heart
1 / 10———
SpireClimb — Slay the Spire 2, Ascension 10
1 / 10———
Safety · 5 benchmarks
MentalHealthBench
6 / 29——2026-11-17
AA-Omniscience Business Non-Hallucination
22 / 26———
AA-Omniscience Law Non-Hallucination
23 / 26———
AA-Omniscience Health Non-Hallucination
24 / 26———
AA-Omniscience v1.1 Hallucination Ratemedium
49 / 76———
Math · 4 benchmarks · first place on 3
OTIS Mock AIME 2024-2025In fit
1 / 141~1.0 yearsJul 2025—
FrontierMath Tier 4 (v2) - PrivateIn fit
1 / 74~3 moFeb 2026—
FrontierMath ErdősIn fit
1 / 16~1 mo~Feb 2028—
FrontierMath Tiers 1-3 (v2) - PrivateIn fit
4 / 91~3 moSep 2026—
Health & Medicine · 3 benchmarks
HealthBench Professional (Length-Adjusted)not reported
2 / 60——2026-11-17
Medical Long Context Reasoning (MLCR-AA, October 2026 methodology)In fit
4 / 114~4 mo~Mar 2029—
HealthBench ProfessionalIn fitnot reported
9 / 40~3 mo~Aug 20282026-11-17
Multimodal · 2 benchmarks · first place on 1
MMMU ProIn fit
1 / 211~4 mo~Apr 2027—
Furniture Assembly BenchmarkIn fit
2 / 40~2 mo~Dec 2026—

GPT-6.2 Astra head to head

Related models

Loading Atlas data…