Astra Faces Its Hallucination Test

Astra’s biggest capability may be knowing when not to answer. OpenAI’s own research acknowledges that evaluation systems can reward guessing over honest uncertainty. [OpenAI research](https://openai.com/index/why-language-models-hallucinate/)

• –BridgeBench seeds tasks with false premises to test whether models correct them or fabricate answers.
• –Developers should evaluate factuality, abstention, tool use, and code execution separately.
• –Lower hallucination rates mean little if achieved through excessive refusals.
• –Astra remains unreleased, so current claims are signals—not production evidence. [OpenAI](https://openai.com/index/responding-next-frontier-critical-cyber-capabilities/)