GPT-6 Astra: benchmark results
Can continue working while tools run and accept corrections mid-task. Reasoning stays enabled even at its lowest effort setting. Provider: OpenAI. Released 2026-09-03. Access: API.
Aggregate Rating 1983 ± 4, rank #4 of 1635 rated models, from 860 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| 4DCodeBench | 0.79 | Reconstruction score (0–1), five-family mean over 200 scenes | 100 |
| AA GPQA Diamond | 96.26 | Accuracy (%) | 100 |
| AA-Omniscience Index - Health | 27 | Omniscience Index | 100 |
| AA-Omniscience Index - Software Engineering (SWE) - Go | 92 | Omniscience Index | 100 |
| AA-Omniscience Index - Software Engineering (SWE) - TypeScript | 97.78 | Omniscience Index | 100 |
| AI for Education Pedagogy | 92.66 | Accuracy (%) | 100 |
| ARC-AGI-3 | 62.71 | Accuracy (%) | 100 |
| AgenticBBO Frontier - BBOB f15 d10 | 0.66 | Normalized anytime/final score (0-1) | 100 |
| AgenticBBO Frontier - Digits MLP HPO | 0.62 | Normalized anytime/final score (0-1) | 100 |
| AgenticBBO Frontier - Sysbench Database Tuning | 0.5 | Normalized anytime/final score (0-1) | 100 |
| BALROG BabaIsAI (LLM) | 100 | Progress (%) | 100 |
| BALROG Crafter (LLM) | 76.82 | Progress (%) | 100 |
Interactive version: theaggregate.ai/model?slug=gpt-6-astra · How It Works · Data refreshed daily, snapshot 2026-10-15.