GPT-6 Astra文档解析基准测试结果
We benchmarked GPT-6 Astra on hard document parsing and extraction tasks. It is the best frontier m...
GPT-6 Astra在文档提取任务上创SOTA，但成本高昂且长文档处理效果差。
GPT-6 Astra在短文档和中等文档的单次提取任务中表现最佳，短文档准确率达97.2%，中等文档达90.6%。该模型在表格理解方面表现优异，但在图表、布局和语义格式处理上存在不足。每页处理成本为11美分，比经济型解决方案高出10倍。
We benchmarked GPT-6 Astra on hard document parsing and extraction tasks. It is the best frontier m...
We benchmarked GPT-6 Astra on hard document parsing and extraction tasks. It is the best frontier model for one-shotting document extraction over short and medium-sized documents: 📊 97.2% over short documents, achieving a new SOTA on our benchmark 📊 90.6% over medium documents But a few caveats: * it's pricy! The average price per page is 11c which is 10x our cost-effective document extraction solution * one-shot performance over long documents is inherently tricky, it only achieves 31.7% * the native OCR capabilities of Astra isn't that much better than other frontier models (e.g. Fable 5.1, gpt-5.6-sol). It does very well on table understanding, but isn't great on charts, layout, and semantic formatting. Full results for document extraction on ExtractBench extractbench.ai c3 💬 3 🔄 1 ❤️ 21 👀 1523 📊 7 ⚡
