GPT-6 Astra (xHigh): benchmark results

Provider: OpenAI. Released 2026-09-03. Access: API.

Aggregate Rating 1968 ± 8, rank #12 of 2257 rated models, from 284 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
AA GPQA Diamond96.26Accuracy (%)100
DeepSWE74.1Pass@1 (%; DeepSWE v1.1 scored attempts)100
DeepSWE-mini (16 Tasks)81.25Pass@1 (%, 16-task subset)100
Epoch AI - Weirdml V342.22Score100
FrontierCS 2 - Go1 Joystick Policy Training50.56Task score (0–100)100
FrontierCS 2 - Latent Objective Recovery99.49Task score (0–100)100
FrontierCS 2 - SQL Counterexample Search97.22Task score (0–100)100
FrontierCS 2 - Sparse Video Attention Policy48.59Task score (0–100)100
LLM2014 Logic 2026-0988.75Median Score100
LLM2014 Logic 2026-1086.53Median Score100
LLM2014 Logic 2026-1186.35Median Score100
NonoBench100Standard accuracy (%), correct / non-failed attempts; versio100

Interactive version: theaggregate.ai/model?slug=gpt-6-astra-xhigh · How It Works · Data refreshed daily, snapshot 2026-11-17.