Similar VIALS-AA scores can result from very different approaches to visual reasoning
Claude Opus 5.5 (max with fallback) averages ~32 tool calls per task, compared with ~8 for GPT-6 Astra, while their VIALS-AA scores are similar (61% and 63%, respectively). This is because Opus performs extensive annotation and measurement, averaging ~8 Image Measurement tool calls per task, compared to ~0.4 for Astra. Image Viewing accounts for a larger share of Astra’s tool calls at ~51% vs. ~42% for Opus