OpenAI GPT-6.2 Astra at launch
- OpenAI
- Released November 17, 2026
- 1.05M context
OpenAI released GPT-6.2 Astra on November 17, 2026. As of November 22, 2026, it ranks #5 of 120 models on the LLMBenchmark Index, between Anthropic Claude Opus 5.5 (#4) and OpenAI GPT-6 Astra (#6). OpenAI's own announcement reports 17 benchmark results, including GPQA Diamond (97.0%) and SEC-Bench Pro (86.8%). It costs $8.00 per million input tokens and $40.00 per million output tokens on OpenRouter.
What OpenAI reported at launch
From OpenAI's announcement, November 17, 2026. The maker's own numbers, not re-run by us.
-
AAV
0.6
Mean Spearman rank correlation (higher is better)
#1 of 9 models with a reported result
-
Coronavirus--ACE2
0.4
Median composite score (higher is better)
#5 of 9 models with a reported result
-
ExploitBench Internal Port (June--August 2026)
53.5%
Mean arbitrary code execution rate across V8 and Chrome
#1 of 5 models with a reported result
-
ExploitBench V8
100.0%
Mean capability credit (five attempts per vulnerability)
#1 of 5 models with a reported result
-
GPQA Diamond
97.0%
no tools; accuracy
#1 of 56 models with a reported result
-
Initial Stage
76.9%
mean rubric score
-
Initial Stage
81.8%
mean rubric score
#1 of 5 models with a reported result
-
MentalHealthBench (non-acute)
53.4
#5 of 10 models with a reported result
-
MentalHealthBench (overall)
53.3
#5 of 10 models with a reported result
-
Multi-select Virology Troubleshooting
61.8%
pass@1; observed
-
Multi-select Virology Troubleshooting
61.9%
pass@1; refusal-adjusted
#2 of 6 models with a reported result
-
Phage--plasmid Co-evolution
11.6
Median of pair-weighted negative log-likelihood (lower is better)
#6 of 9 models with a reported result
Show all 17
-
SEC-Bench Pro
86.8%
Vulnerability reproduction (pass@1)
#3 of 15 models with a reported result
-
Secondary Stage
96.1%
mean reference percentile
#1 of 5 models with a reported result
-
Secondary Stage
42.9%
mean share of baseline training time saved
-
Secondary Stage
74.8%
mean share of performance gap closed
-
SHP2 protein function
0.5
Mean of median R² across three assays (higher is better)
#1 of 5 models with a reported result
Independent results so far
76 scores from 76 boards that test models themselves. Ranks compare GPT-6.2 Astra with every model on the same board.
Agentic work (18)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| Robo-bench (Vals AI) | 32.7% | #1 of 10 | — | Vals AI boards |
| SpireClimb | 41.3% | #1 of 11 | Max | Epoch AI benchmarking hub |
| Terminal-Bench-Science (Vals AI) | 71.4% | #1 of 45 | Mini-SWE-agent | Vals AI boards |
| Vals Science Index | 68.4% | #1 of 27 | — | Vals AI boards |
| Design Arena: agents building web apps | 1334 | #2 of 54 | — | Design Arena |
| BioMysteryBench (Vals AI) | 80.7% | #3 of 33 | — | Vals AI boards |
| Design Arena: agents building full-stack apps | 1351 | #3 of 53 | — | Design Arena |
| Design Arena: agents building mobile apps | 1305 | #3 of 59 | — | Design Arena |
Show all 18
| LiveBench browser | 93.0% | #4 of 35 | — | LiveBench |
|---|---|---|---|---|
| LiveBench research | 91.5% | #4 of 35 | — | LiveBench |
| Design Arena: agents building Android apps | 1257 | #7 of 48 | — | Design Arena |
| LiveBench long agentic | 70.8% | #7 of 35 | — | LiveBench |
| APEX-Agents | 72.2% | #8 of 63 | Max | Epoch AI benchmarking hub |
| Design Arena: spreadsheets | 1278 | #9 of 40 | — | Design Arena |
| Harvey's Legal Agent Benchmark (Vals AI) | 16.3% | #11 of 40 | — | Vals AI boards |
| LMArena Search | 1191 | #17 of 29 | Max | LMArena (Hugging Face dataset) |
| Legal Research Bench (Vals AI) | 42.8% | #28 of 89 | — | Vals AI boards |
| Tax Agent Bench (Vals AI) | 21.8% | #52 of 83 | — | Vals AI boards |
Reasoning (15)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| ARC-AGI | 98.5% | #1 of 77 | Max | Epoch AI benchmarking hub |
| ARC-AGI-2 | 95.8% | #1 of 78 | xHigh | Epoch AI benchmarking hub |
| GPQA Diamond (Epoch AI run) | 96.8% | #1 of 117 | Max | Epoch AI benchmarking hub |
| Mystery game puzzles | 95.0% | #1 of 78 | Max | Epoch AI benchmarking hub |
| MysteryMechanism (Vals AI) | 53.1% | #1 of 32 | — | Vals AI boards |
| EBR-bench | 92.4% | #2 of 32 | Max | Epoch AI benchmarking hub |
| HLE-Diamond (Artificial Analysis run) | 72.2% | #2 of 79 | Max | Artificial Analysis |
| Humanity's Last Exam (Scale run) | 66.9% | #2 of 25 | high | Scale SEAL leaderboards |
Show all 15
| Chess puzzles | 78.0% | #3 of 84 | Max | Epoch AI benchmarking hub |
|---|---|---|---|---|
| CritPt v2, hard subset (Artificial Analysis run) | 71.5% | #3 of 62 | Max | Artificial Analysis |
| GoBench | 2493 | #3 of 20 | High | GoBench (Roland Gao) |
| LiveBench reasoning | 91.6% | #7 of 35 | — | LiveBench |
| SimpleBench | 77.3% | #17 of 78 | — | Epoch AI benchmarking hub |
| LMArena Text: hard prompts | 1425 | #84 of 153 | Max | LMArena (Hugging Face dataset) |
| LMArena Text: hard prompts in English | 1426 | #90 of 153 | Max | LMArena (Hugging Face dataset) |
General (14)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| Furniture assembly | 89.2% | #2 of 39 | Max | Epoch AI benchmarking hub |
| LiveBench data analysis | 84.1% | #4 of 35 | — | LiveBench |
| Vals Index | 67.7% | #8 of 40 | — | Vals Index |
| MultiChallenge | 73.2% | #10 of 28 | high | Scale SEAL leaderboards |
| LMArena Text: writing and language | 1440 | #44 of 153 | Max | LMArena (Hugging Face dataset) |
| LMArena Text: Russian | 1437 | #50 of 153 | Max | LMArena (Hugging Face dataset) |
| LMArena Text: creative writing | 1417 | #52 of 153 | Max | LMArena (Hugging Face dataset) |
| LMArena Text: entertainment and media | 1413 | #55 of 153 | Max | LMArena (Hugging Face dataset) |
Show all 14
| LMArena Text: instruction following | 1423 | #68 of 153 | Max | LMArena (Hugging Face dataset) |
|---|---|---|---|---|
| LMArena Text: longer queries | 1431 | #70 of 153 | Max | LMArena (Hugging Face dataset) |
| LMArena Text: multi-turn | 1425 | #75 of 153 | Max | LMArena (Hugging Face dataset) |
| LMArena Text: non-English | 1404 | #87 of 153 | Max | LMArena (Hugging Face dataset) |
| LMArena Text | 1414 | #89 of 153 | Max | LMArena (Hugging Face dataset) |
| LMArena Text: English | 1417 | #92 of 153 | Max | LMArena (Hugging Face dataset) |
Coding (12)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| LiveBench coding | 91.9% | #1 of 35 | — | LiveBench |
| LMArena WebDev: HTML | 1854 | #1 of 108 | Max | LMArena (Hugging Face dataset) |
| FrontierCode 2.0 | 49.0% | #2 of 62 | xHigh | Artificial Analysis |
| GameBench 2 | 97.2 | #3 of 105 | xHigh | DuelLab GameBench 2 |
| Terminal-Bench 4.0 (Artificial Analysis run) | 64.1% | #3 of 117 | Max | Artificial Analysis |
| LMArena WebDev | 1766 | #5 of 109 | Max | LMArena WebDev |
| Terminal-Bench 4 (Vals AI) | 63.1% | #5 of 63 | Mini-SWE-agent | Vals AI boards |
| MirrorCode | 67.8% | #8 of 38 | Max | Epoch AI benchmarking hub |
Show all 12
| LMArena WebDev: React | 1770 | #9 of 101 | Max | LMArena (Hugging Face dataset) |
|---|---|---|---|---|
| Vibe Code Bench (Vals AI) | 89.9% | #9 of 110 | OpenHands | Vals AI boards |
| LMArena Text: coding | 1450 | #79 of 153 | Max | LMArena (Hugging Face dataset) |
| LMArena Text: software and IT | 1443 | #80 of 153 | Max | LMArena (Hugging Face dataset) |
Web design (6)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| Design Arena: game dev | 1501 | #2 of 134 | — | Design Arena |
| LiveBench design | 82.8% | #3 of 35 | — | LiveBench |
| Design Arena: all code categories | 1389 | #4 of 134 | — | Design Arena |
| Design Arena: 3D | 1430 | #6 of 127 | — | Design Arena |
| Design Arena: websites | 1333 | #7 of 136 | — | Design Arena |
| Design Arena: UI components | 1341 | #10 of 130 | — | Design Arena |
Knowledge (5)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| SimpleQA Verified | 78.5% | #2 of 80 | Max | Epoch AI benchmarking hub |
| AA-Omniscience | 50.0 | #4 of 65 | Max | Artificial Analysis |
| LMArena Text: business and finance | 1413 | #83 of 153 | Max | LMArena (Hugging Face dataset) |
| LMArena Text: legal and government | 1415 | #90 of 151 | Max | LMArena (Hugging Face dataset) |
| LMArena Text: science | 1415 | #97 of 153 | Max | LMArena (Hugging Face dataset) |
Math (4)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| FrontierMath, tier 4 | 100.0% | #1 of 69 | Max | Epoch AI benchmarking hub |
| MathArena | 92.9% | #1 of 73 | max | MathArena |
| OTIS Mock AIME | 100.0% | #1 of 117 | Max | Epoch AI benchmarking hub |
| FrontierMath, tiers 1 to 3 | 92.6% | #4 of 83 | Max | Epoch AI benchmarking hub |
Intelligence (2)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| AA Intelligence Index | 62.1 | #2 of 65 | — | Artificial Analysis (via OpenRouter) |
| Epoch Capabilities Index | 172.8 | #2 of 127 | — | Epoch AI benchmarking hub |