LlamaIndex CEO Reports GPT-6 Astra Extraction Scores

Jerry Liu, co-founder and CEO of LlamaIndex, posted that his team benchmarked GPT-6 Astra on hard document parsing and extraction tasks. He stated the model reached 97.2 percent accuracy on short documents and 90.6 percent on medium documents, describing it as the best frontier model for one-shot extraction and a new state of the art on their benchmark. Liu added that the model is pricey. The benchmark requires every system to fill the same JSON schema from the same enterprise document and scores records on nulls and source boxes.