VentureBeat reports that some frontier large language models can recover up to 65% of facts they cannot directly recall when given more time to reason. The reported finding challenges a common debugging assumption in AI systems: that a factual hallucination necessarily means the model lacks the information internally. According to VentureBeat, developers typically assume hallucinations mean the model lacks the required facts. Engineering teams diagnose the error as missing knowledge, and the standard responses are to increase model size, expand training data, or build retrieval architecture. The reported result points to a different possibility for at least some failures: the fact may be present in the model, but not surfaced on the first attempt. In that framing, longer thinking is not adding new information; it is changing whether the model can recover a fact it could not directly recall. The provided source summary does not name the paper, researchers, benchmark, or specific models behind the 65% figure. That limits how far this summary alone can be generalized. For builders, the practical takeaway is narrow: factual failures may not all mean the model lacks the underlying fact. VentureBeat’s report suggests some failures may be recoverable by giving the model more time to think. Who benefits: AI application teams could benefit if longer reasoning recovers facts that would otherwise be treated as missing knowledge. Who's exposed: Teams that treat every hallucination as missing knowledge may be making an assumption VentureBeat’s report calls into question. The provided source summary does not include methodology details.