A 100-task benchmark built with Ligo finds GPT-6 Astra earns 66.7% of available credit at xhigh effort versus 43.8% for GPT-5.6 Sol, with clear gaps remaining.