GPT-6 Astra System Card

OpenAI's system card for GPT-6 Astra. The companion cites Apollo Research's finding in it that GPT-6 Astra voiced evaluation awareness in 41% of samples, against 28% for GPT-5.5, so low misbehavior rates do not provide substantial evidence about alignment. It also cites the developer's report that the model's monitorability has decreased, supporting the reactive phenotype (P9) and surveillance selection (P10).

Cited in

the companion paper (Companion §8)

Used for

P9P10Prediction 6P9; P10; prediction 6

Strand

AI safety, alignment & autonomy evaluations · System cards & lab safety reports

Cite

@techreport{openai2026astracard,
  title = {GPT-6 Astra System Card},
  author = {OpenAI},
  year = {2026},
  howpublished = {OpenAI Deployment Safety Hub},
  url = {https://deploymentsafety.openai.com/gpt-6-astra},
}