GPT-6 Astra: A new generation of intelligence

OpenAI introduced GPT-6 Astra, saying its test results mark a step change in frontier-model performance. The announcement was published on OpenAI's website under the title 'GPT-6 Astra: A new generation of intelligence.' The post sets out benchmark scores for Astra alongside scores recorded by earlier models. OpenAI presents the comparison as evidence of that step change.

On ExploitBench, OpenAI reports that Astra scored 100%, compared with 78.5% for GPT-5.6 Sol. On ExploitGym, the company reports a 42.4% success rate for Astra, against 30.3% for the prior model. GPT-5.6 Sol is named as the comparison model for ExploitBench, and the ExploitGym figure is described as a success rate. In both cases the reported figure for Astra is the higher of the two, and the ExploitGym figure is below the 100% reported on ExploitBench.

Greg Kamradt of the ARC Prize Foundation said that Astra surpassed its human action-efficiency baseline on 96% of ARC-AGI-3 levels, according to the announcement. The statement concerns ARC-AGI-3 levels and a human action-efficiency baseline, rather than a score or accuracy figure. The comparison is attributed to Kamradt and the ARC Prize Foundation. GPT-6 Astra is described throughout the announcement as a new generation of intelligence, in line with the post's title.