GPT-6.2 Astra beats Opus 5.5 across agent benchmarks at 3-6x lower cost

VraserX · x · 2026-11-18

A side-by-side comparison shows GPT-6.2 Astra beating Claude Opus 5.5: Agents' Last Exam 78% vs 72% ($3 vs $18 per task), Terminal-Bench Science 77% vs 63% ($7 vs $23), and Harvey LAB-AA 15% vs 4% — better results at 3-6x lower cost per task.

Original post →

More from Models

Models channel →