Enter opens the highlighted result. Esc closes.

OpenAI GPT-6.2 Astra

  • OpenAI
  • Released November 17, 2026
  • No public weights listed
  • 1.05M context

As of November 22, 2026, OpenAI GPT-6.2 Astra ranks #5 of 120 models overall with 86.4 ±5.4 on the LLMBenchmark Index. GPT-6.2 Astra ranks #1 of 125 in reasoning, its strongest category. OpenAI GPT-6.2 Astra costs $8.00 per million input tokens and $40.00 per million output tokens. GPT-6.2 Astra was released on November 17, 2026.

GPT-6.2 Astra's output price ranks 333 of 346 priced models, counting from the cheapest.

  • Coding index 87.0 Verified, 4 benchmarks
  • Cost per solved task Not computable No published attempt cost
  • Output, per million tokens $40.00 333 of 346 from the cheapest Maker's API price
  • Input, per million tokens $8.00 Maker's API price
  • Context window 1.05M Tokens per request
New model

GPT-6.2 Astra launch tracker

Listed on OpenRouter November 17, 2026, 5 days before our data. Results so far on 76 of 222 public boards; this page updates every day as boards publish.

About and specs

GPT-6.2 Astra is the flagship model in the GPT-6.2 series from OpenAI, built for the hardest projects. It is suited for real-world engineering, professional work and scientific research, with stronger...

API model id openai/gpt-6.2-astra
  • Tool callingListed
  • Structured outputsListed
  • JSON modeListed
  • Reasoning modeListed
  • Web searchNot listed
Takes in
Files, Images, Text
Gives back
Text
Context window
1.05M tokens
Max output
128K tokens
Knowledge cutoff
May 18, 2026, per OpenAI model documentation
Cached input $/Mtok
$0.40
Cache write $/Mtok
$10.00
Reasoning $/Mtok
Not priced apart
Web search, per request
$0.01
Tokenizer
GPT
Requests moderated
No
Withdrawn on
No date announced
Parameters
Not published
Downloads, all time
No public weights
Training compute
No estimate published
Output speed
45 tokens/s, Max effort
Speed on OpenRouter
43 tokens/s via OpenAI, 30-minute median
First token on OpenRouter
2.6 s via OpenAI, 30-minute median

Benchmarks: 76 of 222 measured

The tick on a coding bar is the median of every model measured on it. Outside coding, the place is this model's rank on that board. A board that never ran the model says so.

Coding index

87.0
  • Terminal-Bench 4 (Vals AI) 63.1% (Mini-SWE-agent) 84 of 100
  • Vibe Code Bench (Vals AI) 89.9% (OpenHands) 90 of 100
  • LiveBench coding 91.9% 92 of 100
  • SWE-bench Verified (Vals AI) Not measured
  • FrontierCode Not measured
  • LMArena WebDev Elo, scaled from 1200 to 1900 1766 Elo (Max) 81 of 100

The index averages only the coding benchmarks that have a result; the others are skipped, never filled in. How it is weighted

Other public boards

All rankings

Best placings

  1. GPQA Diamond (Epoch AI run)1#1 of 117
  2. OTIS Mock AIME1#1 of 117
  3. LMArena WebDev: HTML1#1 of 108
  4. ARC-AGI-21#1 of 78

Lowest placings

  1. LMArena Text: science#97 of 153
  2. Tax Agent Bench (Vals AI)#52 of 83
  3. LMArena Text: English#92 of 153

Coding8 of 46 boards

2 more coding boards

Not measured on Aider polyglot, SWE-bench Verified, SWE-bench Lite, SWE-bench (full test set), SWE-bench Multimodal, SWE-bench Multilingual, AA Coding Index, LiveCodeBench (Vals AI), Code migration (Vals AI), ProgramBench (Vals AI), IOI (Vals AI), Terminal-Bench 2.1 (Vals AI), SWE-bench Multimodal (Vals AI), Vibe Code Bench 1-100 (Vals AI), SkillsBench (Vals AI), CyberBench (Vals AI), LiveCodeBench (Artificial Analysis run), SciCode, Terminal-Bench Hard (Artificial Analysis run), Terminal-Bench, DeepSWE, FrontierSWE, GSO-Bench, WeirdML, Cybench, ALE-Bench, LMArena WebDev: from an image, SciCode, SWE-Bench Pro V2, SWE Atlas: Codebase QnA, SWE Atlas: Refactoring, SWE Atlas: Test Writing, Terminal-Bench 2.0, Terminal-Bench 2.1, Terminal-Bench 3.0, Terminal-Bench 4.0, CursorBench 4.0, ProgramBench

Intelligence2 of 2 boards

Agentic work18 of 44 boards

12 more agentic work boards

Not measured on AA Agentic Index, Design Arena: agents building Godot games, Design Arena: agents building games, Design Arena: HTML slides, Design Arena: PowerPoint slides, SREBench (Vals AI), Finance Agent (Vals AI), Excel Modeling Benchmark (Vals AI), Time Horizon Index (Vals AI), CUA-bench (Vals AI), τ²-Bench Telecom, METR time horizons, Remote Labor Index, GDPval, OSWorld 2.0, BALROG, DeepResearch Bench, PostTrainBench, TheAgentCompany, OSWorld, Vending-Bench 2, Berkeley Function Calling Leaderboard (BFCL), BrowseComp-Plus, MCP Atlas, Toolathlon-Verified, τ³-Banking

Reasoning15 of 26 boards

9 more reasoning boards

Not measured on GPQA Diamond, MMMU Pro (Vals AI), GPQA Diamond (Artificial Analysis run), Humanity's Last Exam (Artificial Analysis run), Humanity's Last Exam, VPCT, BIG-Bench Hard, LMArena Text: expert, LMArena Vision: diagrams, LMArena Vision: homework, Kagi LLM Benchmark

Math4 of 14 boards

Not measured on AIME, MATH 500 (Vals AI), MGSM (Vals AI), AA Math Index, AIME 2025 (Artificial Analysis run), ProofBench, MATH level 5, GSM8K, LMArena Text: math, LMArena Text: mathematical work

Knowledge5 of 21 boards

Not measured on MMLU Pro, LegalBench (Vals AI), CaseLaw (Vals AI), CorpFin (Vals AI), TaxEval (Vals AI), MortgageTax (Vals AI), MedQA (Vals AI), MedCode (Vals AI), MMLU Pro (Artificial Analysis run), MMLU, GeoBench, AI2 Reasoning Challenge, LMArena Text: medicine and healthcare, FACTS Benchmark Suite, PRBench Finance, PRBench Legal

Web design6 of 9 boards

Not measured on Design Arena: data viz, Design Arena: SVG, Design Arena: ASCII art

General14 of 49 boards

8 more general boards

Not measured on Public Benefits Bench (Vals AI), SAGE (Vals AI), MedScribe (Vals AI), Vals Multimodal Index, IFBench, Long context reasoning (AA-LCR), GDPval-AA, DTBench, LMCA, Fiction.LiveBench, Creative writing (Lech Mazur), CL-bench, Surface Evolver Bench, ExploitBench, HellaSwag, WinoGrande, PIQA, CL-bench Life, LMArena Vision, LMArena Text: Chinese, LMArena Text: German, LMArena Text: Japanese, LMArena Text: Korean, LMArena Text: French, LMArena Text: Spanish, LMArena Text: Polish, LMArena Vision: English, LMArena Vision: Chinese, LMArena Vision: reading text, LMArena Vision: humor, LMArena Vision: captioning, LMArena Vision: recognizing things, LMArena Vision: creative writing, LMArena Document, EQ-Bench 4

Image generation0 of 5 boards

Not measured on Design Arena: images, Design Arena: graphic design, Design Arena: logos, LMArena Text to Image, LMArena Image Edit

Public boards shown as published (Vals AI, Artificial Analysis, Design Arena) and kept out of the coding index. Each links to its source on the benchmark page.

All 222 benchmarks as a table
Benchmark results for OpenAI GPT-6.2 Astra
BenchmarkResultIndex weight
Aider polyglotCoding Not measured Not in index
SWE-bench VerifiedCoding Not measured Not in index
SWE-bench LiteCoding Not measured Not in index
SWE-bench (full test set)Coding Not measured Not in index
SWE-bench MultimodalCoding Not measured Not in index
SWE-bench MultilingualCoding Not measured Not in index
LiveBench codingCoding 91.9% 15%
LMArena WebDevCoding 1766(Max) 10%
Vals IndexGeneral 67.7% Not in index
GPQA DiamondReasoning Not measured Not in index
AIMEMath Not measured Not in index
MMLU ProKnowledge Not measured Not in index
AA Intelligence IndexIntelligence 62.1 Not in index
AA Coding IndexCoding Not measured Not in index
AA Agentic IndexAgentic work Not measured Not in index
Design Arena: websitesWeb design 1333 Not in index
Design Arena: UI componentsWeb design 1341 Not in index
Design Arena: game devWeb design 1501 Not in index
Design Arena: data vizWeb design Not measured Not in index
Design Arena: 3DWeb design 1430 Not in index
Design Arena: all code categoriesWeb design 1389 Not in index
Design Arena: SVGWeb design Not measured Not in index
Design Arena: ASCII artWeb design Not measured Not in index
Design Arena: agents building web appsAgentic work 1334 Not in index
Design Arena: agents building full-stack appsAgentic work 1351 Not in index
Design Arena: agents building mobile appsAgentic work 1305 Not in index
Design Arena: agents building Android appsAgentic work 1257 Not in index
Design Arena: agents building Godot gamesAgentic work Not measured Not in index
Design Arena: agents building gamesAgentic work Not measured Not in index
Design Arena: HTML slidesAgentic work Not measured Not in index
Design Arena: PowerPoint slidesAgentic work Not measured Not in index
Design Arena: spreadsheetsAgentic work 1278 Not in index
Design Arena: imagesImage generation Not measured Not in index
Design Arena: graphic designImage generation Not measured Not in index
Design Arena: logosImage generation Not measured Not in index
LiveBench reasoningReasoning 91.6% Not in index
LiveBench data analysisGeneral 84.1% Not in index
LiveBench researchAgentic work 91.5% Not in index
LiveBench browserAgentic work 93.0% Not in index
LiveBench long agenticAgentic work 70.8% Not in index
LiveBench designWeb design 82.8% Not in index
LiveCodeBench (Vals AI)Coding Not measured Not in index
SWE-bench Verified (Vals AI)Coding Not measured 15%
Terminal-Bench 4 (Vals AI)Coding 63.1%(Mini-SWE-agent) 25%
Vibe Code Bench (Vals AI)Coding 89.9%(OpenHands) 20%
Code migration (Vals AI)Coding Not measured Not in index
ProgramBench (Vals AI)Coding Not measured Not in index
SREBench (Vals AI)Agentic work Not measured Not in index
IOI (Vals AI)Coding Not measured Not in index
Terminal-Bench 2.1 (Vals AI)Coding Not measured Not in index
SWE-bench Multimodal (Vals AI)Coding Not measured Not in index
Vibe Code Bench 1-100 (Vals AI)Coding Not measured Not in index
SkillsBench (Vals AI)Coding Not measured Not in index
CyberBench (Vals AI)Coding Not measured Not in index
Finance Agent (Vals AI)Agentic work Not measured Not in index
Tax Agent Bench (Vals AI)Agentic work 21.8% Not in index
Legal Research Bench (Vals AI)Agentic work 42.8% Not in index
Harvey's Legal Agent Benchmark (Vals AI)Agentic work 16.3% Not in index
Excel Modeling Benchmark (Vals AI)Agentic work Not measured Not in index
BioMysteryBench (Vals AI)Agentic work 80.7% Not in index
Terminal-Bench-Science (Vals AI)Agentic work 71.4%(Mini-SWE-agent) Not in index
Vals Science IndexAgentic work 68.4% Not in index
Time Horizon Index (Vals AI)Agentic work Not measured Not in index
CUA-bench (Vals AI)Agentic work Not measured Not in index
Robo-bench (Vals AI)Agentic work 32.7% Not in index
Public Benefits Bench (Vals AI)General Not measured Not in index
MysteryMechanism (Vals AI)Reasoning 53.1% Not in index
MMMU Pro (Vals AI)Reasoning Not measured Not in index
MATH 500 (Vals AI)Math Not measured Not in index
MGSM (Vals AI)Math Not measured Not in index
SAGE (Vals AI)General Not measured Not in index
LegalBench (Vals AI)Knowledge Not measured Not in index
CaseLaw (Vals AI)Knowledge Not measured Not in index
CorpFin (Vals AI)Knowledge Not measured Not in index
TaxEval (Vals AI)Knowledge Not measured Not in index
MortgageTax (Vals AI)Knowledge Not measured Not in index
MedQA (Vals AI)Knowledge Not measured Not in index
MedCode (Vals AI)Knowledge Not measured Not in index
MedScribe (Vals AI)General Not measured Not in index
Vals Multimodal IndexGeneral Not measured Not in index
AA Math IndexMath Not measured Not in index
GPQA Diamond (Artificial Analysis run)Reasoning Not measured Not in index
Humanity's Last Exam (Artificial Analysis run)Reasoning Not measured Not in index
MMLU Pro (Artificial Analysis run)Knowledge Not measured Not in index
LiveCodeBench (Artificial Analysis run)Coding Not measured Not in index
SciCodeCoding Not measured Not in index
Terminal-Bench Hard (Artificial Analysis run)Coding Not measured Not in index
IFBenchGeneral Not measured Not in index
Long context reasoning (AA-LCR)General Not measured Not in index
τ²-Bench TelecomAgentic work Not measured Not in index
AIME 2025 (Artificial Analysis run)Math Not measured Not in index
Terminal-Bench 4.0 (Artificial Analysis run)Coding 64.1%(Max) Not in index
FrontierCode 2.0Coding 49.0%(xHigh) Not in index
HLE-Diamond (Artificial Analysis run)Reasoning 72.2%(Max) Not in index
GDPval-AAGeneral Not measured Not in index
AA-OmniscienceKnowledge 50.0(Max) Not in index
CritPt v2, hard subset (Artificial Analysis run)Reasoning 71.5%(Max) Not in index
Terminal-BenchCoding Not measured Not in index
DeepSWECoding Not measured Not in index
FrontierCodeCoding Not measured 15%
FrontierSWECoding Not measured Not in index
GSO-BenchCoding Not measured Not in index
WeirdMLCoding Not measured Not in index
MirrorCodeCoding 67.8%(Max) Not in index
APEX-AgentsAgentic work 72.2%(Max) Not in index
METR time horizonsAgentic work Not measured Not in index
Remote Labor IndexAgentic work Not measured Not in index
GDPvalAgentic work Not measured Not in index
OSWorld 2.0Agentic work Not measured Not in index
BALROGAgentic work Not measured Not in index
DeepResearch BenchAgentic work Not measured Not in index
PostTrainBenchAgentic work Not measured Not in index
GPQA Diamond (Epoch AI run)Reasoning 96.8%(Max) Not in index
ARC-AGIReasoning 98.5%(Max) Not in index
ARC-AGI-2Reasoning 95.8%(xHigh) Not in index
Humanity's Last ExamReasoning Not measured Not in index
SimpleBenchReasoning 77.3% Not in index
Chess puzzlesReasoning 78.0%(Max) Not in index
Mystery game puzzlesReasoning 95.0%(Max) Not in index
EBR-benchReasoning 92.4%(Max) Not in index
OTIS Mock AIMEMath 100.0%(Max) Not in index
FrontierMath, tiers 1 to 3Math 92.6%(Max) Not in index
FrontierMath, tier 4Math 100.0%(Max) Not in index
ProofBenchMath Not measured Not in index
MATH level 5Math Not measured Not in index
GSM8KMath Not measured Not in index
SimpleQA VerifiedKnowledge 78.5%(Max) Not in index
MMLUKnowledge Not measured Not in index
GeoBenchKnowledge Not measured Not in index
DTBenchGeneral Not measured Not in index
LMCAGeneral Not measured Not in index
Fiction.LiveBenchGeneral Not measured Not in index
Creative writing (Lech Mazur)General Not measured Not in index
CL-benchGeneral Not measured Not in index
Surface Evolver BenchGeneral Not measured Not in index
Furniture assemblyGeneral 89.2%(Max) Not in index
ExploitBenchGeneral Not measured Not in index
CybenchCoding Not measured Not in index
TheAgentCompanyAgentic work Not measured Not in index
OSWorldAgentic work Not measured Not in index
SpireClimbAgentic work 41.3%(Max) Not in index
VPCTReasoning Not measured Not in index
BIG-Bench HardReasoning Not measured Not in index
AI2 Reasoning ChallengeKnowledge Not measured Not in index
HellaSwagGeneral Not measured Not in index
WinoGrandeGeneral Not measured Not in index
PIQAGeneral Not measured Not in index
CL-bench LifeGeneral Not measured Not in index
ALE-BenchCoding Not measured Not in index
Vending-Bench 2Agentic work Not measured Not in index
Epoch Capabilities IndexIntelligence 172.8 Not in index
LMArena TextGeneral 1414(Max) Not in index
LMArena Text: codingCoding 1450(Max) Not in index
LMArena Text: hard promptsReasoning 1425(Max) Not in index
LMArena Text: mathMath Not measured Not in index
LMArena Text: instruction followingGeneral 1423(Max) Not in index
LMArena Text: creative writingGeneral 1417(Max) Not in index
LMArena VisionGeneral Not measured Not in index
LMArena Text: EnglishGeneral 1417(Max) Not in index
LMArena Text: non-EnglishGeneral 1404(Max) Not in index
LMArena Text: ChineseGeneral Not measured Not in index
LMArena Text: GermanGeneral Not measured Not in index
LMArena Text: JapaneseGeneral Not measured Not in index
LMArena Text: KoreanGeneral Not measured Not in index
LMArena Text: FrenchGeneral Not measured Not in index
LMArena Text: SpanishGeneral Not measured Not in index
LMArena Text: RussianGeneral 1437(Max) Not in index
LMArena Text: PolishGeneral Not measured Not in index
LMArena Text: multi-turnGeneral 1425(Max) Not in index
LMArena Text: longer queriesGeneral 1431(Max) Not in index
LMArena Text: expertReasoning Not measured Not in index
LMArena Text: hard prompts in EnglishReasoning 1426(Max) Not in index
LMArena Text: software and ITCoding 1443(Max) Not in index
LMArena Text: mathematical workMath Not measured Not in index
LMArena Text: scienceKnowledge 1415(Max) Not in index
LMArena Text: medicine and healthcareKnowledge Not measured Not in index
LMArena Text: legal and governmentKnowledge 1415(Max) Not in index
LMArena Text: business and financeKnowledge 1413(Max) Not in index
LMArena Text: writing and languageGeneral 1440(Max) Not in index
LMArena Text: entertainment and mediaGeneral 1413(Max) Not in index
LMArena WebDev: ReactCoding 1770(Max) Not in index
LMArena WebDev: HTMLCoding 1854(Max) Not in index
LMArena WebDev: from an imageCoding Not measured Not in index
LMArena Vision: EnglishGeneral Not measured Not in index
LMArena Vision: ChineseGeneral Not measured Not in index
LMArena Vision: reading textGeneral Not measured Not in index
LMArena Vision: diagramsReasoning Not measured Not in index
LMArena Vision: homeworkReasoning Not measured Not in index
LMArena Vision: humorGeneral Not measured Not in index
LMArena Vision: captioningGeneral Not measured Not in index
LMArena Vision: recognizing thingsGeneral Not measured Not in index
LMArena Vision: creative writingGeneral Not measured Not in index
LMArena SearchAgentic work 1191(Max) Not in index
LMArena DocumentGeneral Not measured Not in index
LMArena Text to ImageImage generation Not measured Not in index
LMArena Image EditImage generation Not measured Not in index
Berkeley Function Calling Leaderboard (BFCL)Agentic work Not measured Not in index
Kagi LLM BenchmarkReasoning Not measured Not in index
BrowseComp-PlusAgentic work Not measured Not in index
SciCodeCoding Not measured Not in index
FACTS Benchmark SuiteKnowledge Not measured Not in index
MathArenaMath 92.9%(max) Not in index
MultiChallengeGeneral 73.2%(high) Not in index
MCP AtlasAgentic work Not measured Not in index
Humanity's Last Exam (Scale run)Reasoning 66.9%(high) Not in index
PRBench FinanceKnowledge Not measured Not in index
PRBench LegalKnowledge Not measured Not in index
SWE-Bench Pro V2Coding Not measured Not in index
SWE Atlas: Codebase QnACoding Not measured Not in index
SWE Atlas: RefactoringCoding Not measured Not in index
SWE Atlas: Test WritingCoding Not measured Not in index
EQ-Bench 4General Not measured Not in index
GameBench 2Coding 97.2(xHigh) Not in index
GoBenchReasoning 2493(High) Not in index
Terminal-Bench 2.0Coding Not measured Not in index
Terminal-Bench 2.1Coding Not measured Not in index
Terminal-Bench 3.0Coding Not measured Not in index
Terminal-Bench 4.0Coding Not measured Not in index
CursorBench 4.0Coding Not measured Not in index
Toolathlon-VerifiedAgentic work Not measured Not in index
ProgramBenchCoding Not measured Not in index
τ³-BankingAgentic work Not measured Not in index

Scores by reasoning effort

Some boards test this model at more than one reasoning effort. Rankings use its best run; every run the source published is listed here.

Scores by reasoning effort
BenchmarkEffortScoreSource
AA-Omniscience Max · used in rankings 50.0 Artificial Analysis
AA-Omniscience xHigh 48.3 Artificial Analysis
AA-Omniscience High 46.8 Artificial Analysis
AA-Omniscience Medium 44.8 Artificial Analysis
ARC-AGI Max · used in rankings 98.5% Epoch AI benchmarking hub
ARC-AGI High 98.5% Epoch AI benchmarking hub
ARC-AGI xHigh 98.5% Epoch AI benchmarking hub
ARC-AGI Low 98.5% Epoch AI benchmarking hub
ARC-AGI Medium 98.5% Epoch AI benchmarking hub
ARC-AGI-2 xHigh · used in rankings 95.8% Epoch AI benchmarking hub
ARC-AGI-2 Max 95.8% Epoch AI benchmarking hub
ARC-AGI-2 High 94.6% Epoch AI benchmarking hub
ARC-AGI-2 Medium 93.3% Epoch AI benchmarking hub
ARC-AGI-2 Low 91.7% Epoch AI benchmarking hub
CritPt v2, hard subset (Artificial Analysis run) Max · used in rankings 71.5% Artificial Analysis
CritPt v2, hard subset (Artificial Analysis run) High 71.5% Artificial Analysis
CritPt v2, hard subset (Artificial Analysis run) xHigh 70.5% Artificial Analysis
CritPt v2, hard subset (Artificial Analysis run) Medium 67.5% Artificial Analysis
FrontierCode 2.0 xHigh · used in rankings 49.0% Artificial Analysis
FrontierCode 2.0 High 47.1% Artificial Analysis
FrontierCode 2.0 Medium 46.6% Artificial Analysis
FrontierCode 2.0 Max 45.4% Artificial Analysis
GameBench 2 xHigh · used in rankings 97.2 DuelLab GameBench 2
GameBench 2 Medium 92.9 DuelLab GameBench 2
GameBench 2 Low 90.7 DuelLab GameBench 2
GoBench High · used in rankings 2493 GoBench (Roland Gao)
GoBench xHigh 2363 GoBench (Roland Gao)
HLE-Diamond (Artificial Analysis run) Max · used in rankings 72.2% Artificial Analysis
HLE-Diamond (Artificial Analysis run) xHigh 67.3% Artificial Analysis
HLE-Diamond (Artificial Analysis run) High 65.9% Artificial Analysis
HLE-Diamond (Artificial Analysis run) Medium 56.0% Artificial Analysis
Terminal-Bench 4.0 (Artificial Analysis run) Max · used in rankings 64.1% Artificial Analysis
Terminal-Bench 4.0 (Artificial Analysis run) xHigh 64.1% Artificial Analysis
Terminal-Bench 4.0 (Artificial Analysis run) High 60.1% Artificial Analysis
Terminal-Bench 4.0 (Artificial Analysis run) Medium 58.6% Artificial Analysis

What it costs

A list price says what a token costs. The cost per solved task says what a working answer costs, which is the number that decides.

Output price against the market

GPT-6.2 Astra pricing
$0.045Dashed line: median $2.00, log scale$180.00
$40.00 per million output tokens is cheaper than 3% of the 341 paid models in the catalog.

Cost per solved task

How it is computed

Neither board that publishes what a SWE-bench Verified attempt costs, Vals AI and swebench.com, has run this model. We leave the number out rather than estimate it.

LiveBench, on its own coding tasks, publishes what one task cost this model: $5.122, at a score of 91.9%. LiveBench counts that as $5.574 per successful task. Different tasks from SWE-bench, so it is not our number and does not enter its ranking. LiveBench coding

Nearly as good as GPT-6.2 Astra, for less

What the boards say about its runs

Beside a result, many boards publish what the run cost, how long it took, the tokens it spent or the votes behind a rating. Each figure as its board printed it, one board per row.

What each board publishes about the runs behind OpenAI: GPT-6.2 Astra's results
BoardResultWhat the run costAlso published
LiveBench coding 91.9% $5.1222 per task Output tokens per task 36,763
Vals Index 67.7% $10.68 per test Nothing else
LiveBench reasoning 91.6% $0.1323 per task Output tokens per task 3,049
LiveBench data analysis 84.1% $0.3547 per task Output tokens per task 7,950
LiveBench research 91.5% $6.6314 per task Output tokens per task 29,807
LiveBench browser 93.0% $2.5049 per task Output tokens per task 5,611
LiveBench long agentic 70.8% $17.0234 per task Output tokens per task 123,129
LiveBench design 82.8% $1.9926 per task Output tokens per task 30,037
Terminal-Bench 4 (Vals AI) 63.1%(Mini-SWE-agent) $4.69 per test Nothing else
Vibe Code Bench (Vals AI) 89.9%(OpenHands) $7.92 per test Nothing else
Tax Agent Bench (Vals AI) 21.8% $7.90 per test Nothing else
Legal Research Bench (Vals AI) 42.8% $12.84 per test Nothing else
40 more boards
What each board publishes about the runs behind OpenAI: GPT-6.2 Astra's results, continued
Harvey's Legal Agent Benchmark (Vals AI) 16.3% $7.33 per test Nothing else
BioMysteryBench (Vals AI) 80.7% $1.49 per test Nothing else
Terminal-Bench-Science (Vals AI) 71.4%(Mini-SWE-agent) $11.08 per test Nothing else
Vals Science Index 68.4% $4.76 per test Nothing else
MysteryMechanism (Vals AI) 53.1% $1.72 per test Nothing else
MathArena 92.9%(max) $2.4633 expected cost per problem Nothing else
GameBench 2 97.2(xHigh) $0.6788 per program First-try builds 100.0% · Failed programs 0.0% · Matches 5,358 · Wins 5,076 · Draws 106 · Losses 176
GoBench 2493 Elo(High) $0.024 per move Interval (±) ±136.0 · Seconds per move 17 s
LMArena WebDev 1766 Elo(Max) Not published Votes 1,913 · Interval (±) ±15.3
Design Arena: websites 1333 Elo Not published Battles 887
Design Arena: UI components 1341 Elo Not published Battles 139
Design Arena: game dev 1501 Elo Not published Battles 588
Design Arena: 3D 1430 Elo Not published Battles 340
Design Arena: all code categories 1389 Elo Not published Battles 1,801
Design Arena: agents building web apps 1334 Elo Not published Battles 597
Design Arena: agents building full-stack apps 1351 Elo Not published Battles 343
Design Arena: agents building mobile apps 1305 Elo Not published Battles 194
Design Arena: agents building Android apps 1257 Elo Not published Battles 173
Design Arena: spreadsheets 1278 Elo Not published Battles 32
APEX-Agents 72.2%(Max) Not published Mean score 83.9%
LMArena Text 1414 Elo(Max) Not published Votes 2,398 · Interval (±) ±12.4
LMArena Text: coding 1450 Elo(Max) Not published Votes 473 · Interval (±) ±27.3
LMArena Text: hard prompts 1425 Elo(Max) Not published Votes 1,439 · Interval (±) ±15.8
LMArena Text: instruction following 1423 Elo(Max) Not published Votes 791 · Interval (±) ±20.8
LMArena Text: creative writing 1417 Elo(Max) Not published Votes 413 · Interval (±) ±30.3
LMArena Text: English 1417 Elo(Max) Not published Votes 749 · Interval (±) ±21.6
LMArena Text: non-English 1404 Elo(Max) Not published Votes 1,635 · Interval (±) ±14.8
LMArena Text: Russian 1437 Elo(Max) Not published Votes 472 · Interval (±) ±26.7
LMArena Text: multi-turn 1425 Elo(Max) Not published Votes 329 · Interval (±) ±32.6
LMArena Text: longer queries 1431 Elo(Max) Not published Votes 977 · Interval (±) ±19.4
LMArena Text: hard prompts in English 1426 Elo(Max) Not published Votes 447 · Interval (±) ±27.9
LMArena Text: software and IT 1443 Elo(Max) Not published Votes 802 · Interval (±) ±20.9
LMArena Text: science 1415 Elo(Max) Not published Votes 384 · Interval (±) ±29.7
LMArena Text: legal and government 1415 Elo(Max) Not published Votes 183 · Interval (±) ±40.9
LMArena Text: business and finance 1413 Elo(Max) Not published Votes 379 · Interval (±) ±29.8
LMArena Text: writing and language 1440 Elo(Max) Not published Votes 620 · Interval (±) ±24.1
LMArena Text: entertainment and media 1413 Elo(Max) Not published Votes 542 · Interval (±) ±26.1
LMArena WebDev: React 1770 Elo(Max) Not published Votes 1,221 · Interval (±) ±19.0
LMArena WebDev: HTML 1854 Elo(Max) Not published Votes 302 · Interval (±) ±44.8
LMArena Search 1191 Elo(Max) Not published Votes 1,077 · Interval (±) ±20.4

Results OpenAI reported at launch

OpenAI published these results with the model's release on November 17, 2026. They are the maker's own numbers, not independent measurements, and they do not count toward the LLMBenchmark Index.

Results OpenAI reported for this model
BenchmarkScoreSettingSource
AAV 0.6 Mean Spearman rank correlation (higher is better) November 17, 2026 (model release)
Coronavirus--ACE2 0.4 Median composite score (higher is better) November 17, 2026 (model release)
ExploitBench Internal Port (June--August 2026) 53.5% Mean arbitrary code execution rate across V8 and Chrome November 17, 2026 (model release)
ExploitBench V8 100.0% Mean capability credit (five attempts per vulnerability) November 17, 2026 (model release)
GPQA Diamond 97.0% no tools; accuracy November 17, 2026 (model release)
Initial Stage 81.8% mean rubric score November 17, 2026 (model release)
Initial Stage 76.9% mean rubric score November 17, 2026 (model release)
MentalHealthBench (non-acute) 53.4 Not published November 17, 2026 (model release)
MentalHealthBench (overall) 53.3 Not published November 17, 2026 (model release)
Multi-select Virology Troubleshooting 61.8% pass@1; observed November 17, 2026 (model release)
Multi-select Virology Troubleshooting 61.9% pass@1; refusal-adjusted November 17, 2026 (model release)
Phage--plasmid Co-evolution 11.6 Median of pair-weighted negative log-likelihood (lower is better) November 17, 2026 (model release)
SEC-Bench Pro 86.8% Vulnerability reproduction (pass@1) November 17, 2026 (model release)
Secondary Stage 96.1% mean reference percentile November 17, 2026 (model release)
Secondary Stage 42.9% mean share of baseline training time saved November 17, 2026 (model release)
Secondary Stage 74.8% mean share of performance gap closed November 17, 2026 (model release)
SHP2 protein function 0.5 Mean of median R² across three assays (higher is better) November 17, 2026 (model release)

Reported later, in other OpenAI announcements

OpenAI also published numbers for this model in later announcements, usually as a comparison with a newer model. They are listed apart because they come from a different release.

Results OpenAI reported for this model in later announcements
BenchmarkScoreSettingSource
AAV 0.6 Mean Spearman rank correlation (higher is better) November 17, 2026 (model release)
AAV 0.6 Mean Spearman rank correlation (higher is better) November 17, 2026 (model release)
Coronavirus--ACE2 0.4 Median composite score (higher is better) November 17, 2026 (model release)
Coronavirus--ACE2 0.4 Median composite score (higher is better) November 17, 2026 (model release)
ExploitBench Internal Port (June--August 2026) 53.5% Mean arbitrary code execution rate across V8 and Chrome November 17, 2026 (model release)
ExploitBench Internal Port (June--August 2026) 53.5% Mean arbitrary code execution rate across V8 and Chrome November 17, 2026 (model release)
ExploitBench V8 100.0% Mean capability credit (five attempts per vulnerability) November 17, 2026 (model release)
ExploitBench V8 100.0% Mean capability credit (five attempts per vulnerability) November 17, 2026 (model release)
GPQA Diamond 97.0% no tools; accuracy November 17, 2026 (model release)
GPQA Diamond 97.0% no tools; accuracy November 17, 2026 (model release)
Initial Stage 81.8% mean rubric score November 17, 2026 (model release)
Initial Stage 76.9% mean rubric score November 17, 2026 (model release)
Initial Stage 81.8% mean rubric score November 17, 2026 (model release)
Initial Stage 76.9% mean rubric score November 17, 2026 (model release)
MentalHealthBench (emergent) 51 Not published November 17, 2026 (model release)
MentalHealthBench (emergent) 51 Not published November 17, 2026 (model release)
MentalHealthBench (high acuity) 56.4 high-acuity November 17, 2026 (model release)
MentalHealthBench (high acuity) 56.4 high-acuity November 17, 2026 (model release)
MentalHealthBench (non-acute) 53.4 Not published November 17, 2026 (model release)
MentalHealthBench (non-acute) 53.4 Not published November 17, 2026 (model release)
MentalHealthBench (overall) 53.3 Not published November 17, 2026 (model release)
MentalHealthBench (overall) 53.3 Not published November 17, 2026 (model release)
Multi-select Virology Troubleshooting 61.8% pass@1; observed November 17, 2026 (model release)
Multi-select Virology Troubleshooting 61.8% pass@1; observed November 17, 2026 (model release)
Multi-select Virology Troubleshooting 61.9% pass@1; refusal-adjusted November 17, 2026 (model release)
Multi-select Virology Troubleshooting 61.9% pass@1; refusal-adjusted November 17, 2026 (model release)
Phage--plasmid Co-evolution 11.6 Median of pair-weighted negative log-likelihood (lower is better) November 17, 2026 (model release)
Phage--plasmid Co-evolution 11.6 Median of pair-weighted negative log-likelihood (lower is better) November 17, 2026 (model release)
SEC-Bench Pro 86.8% Vulnerability reproduction (pass@1) November 17, 2026 (model release)
SEC-Bench Pro 86.8% Vulnerability reproduction (pass@1) November 17, 2026 (model release)
Secondary Stage 96.1% mean reference percentile November 17, 2026 (model release)
Secondary Stage 96.1% mean reference percentile November 17, 2026 (model release)
Secondary Stage 42.9% mean share of baseline training time saved November 17, 2026 (model release)
Secondary Stage 42.9% mean share of baseline training time saved November 17, 2026 (model release)
Secondary Stage 74.8% mean share of performance gap closed November 17, 2026 (model release)
Secondary Stage 74.8% mean share of performance gap closed November 17, 2026 (model release)
SHP2 protein function 0.5 Mean of median R² across three assays (higher is better) November 17, 2026 (model release)
SHP2 protein function 0.5 Mean of median R² across three assays (higher is better) November 17, 2026 (model release)

OpenAI GPT-6.2 Astra: launch results and current rankings

OpenAI GPT-6.2 Astra API providers

On these lists

Compare with models of a similar score

All 110 comparisons for GPT-6.2 Astra